Washington State
Institute for
Public Policy
110 Fifth Avenue Southeast, Suite 214 • PO Box 40999 • Olympia, WA 98504-0999 • (360) 586-2677 • FAX (360) 586-2793 • www.wsipp.wa.gov
Initial Report
BENEFITS AND COSTS OF K–12 EDUCATIONAL POLICIES:
Evidence-Based Effects of Class Size Reductions and Full-Day Kindergarten
The Washington State Legislature directed the
Washington State Institute for Public Policy (Institute)
“to begin the development of a repository of research
and evaluations of the cost-benefits of various K–12
educational programs and services.”1
This report contains our initial findings on two
topics: class size reductions and full-day
kindergarten. We examine existing research
evidence to estimate whether student academic
achievement can be expected to improve with each
policy. We also compute the expected return on
investment for the two options. Upcoming reports
will include other K–12 topics.
This research assignment from the Legislature is
designed to augment the recent Washington Learns
process—the statewide effort to identify ways to
improve Washington’s early learning, K–12, and
higher education systems. In its final report issued
in November 2006, the Washington Learns
Steering Committee adopted principles for
changing Washington’s public education system.
Among other recommendations, the Committee
stated that Washington “will invest only in programs
that work” and that the state “must be diligent about
redirecting current educational dollars into proven
strategies for improved results.”2
Following these principles, the purpose of this
research is to estimate the likely costs and benefits
of “research-proven” K–12 policies and programs.
Any attempt to calculate costs and benefits
encounters a high analytical bar. Conducting this
type of study implies being able to answer central
questions about causality. That is, if costs are
incurred, will benefits be obtained?
Continues on page 2
‡
Suggested citation: Steve Aos, Marna Miller, & Jim Mayfield.
(2007). Benefits and Costs of K–12 Educational Policies:
Evidence-Based Effects of Class Size Reductions and Full-
Day Kindergarten. Olympia: Washington State Institute for
Public Policy, Document No. 07-03-2201.
Contact email: saos@wsipp.wa.gov.
March 2007
‡
Summary
The Washington Legislature directed the
Washington State Institute for Public Policy to
begin conducting economic analyses of certain
K–12 policies. Augmenting the work of the recent
Washington Learns process, this report describes
our initial cost-benefit findings for class size
reductions and full-day vs. half-day kindergarten.
Upcoming reports will examine other K–12 topics.
Research Approach
We examine all rigorous research studies to
estimate whether academic achievement can be
expected to improve with each policy. We also
compute an expected return on investment by
estimating long-run labor market and other non-
market benefits of improved academic outcomes.
Finding: Class Size Reductions
We analyze 38 recent high-quality evaluations of
whether reducing the number of students in a
classroom improves student test scores. The
results are mixed. We find that during kindergarten
through second grade, there is evidence that
reducing class size increases test scores. During
third through sixth grade, the gains remain
significant but are much smaller—only 35 percent
of the kindergarten through second grade gains. In
middle and high school, we find that reduced class
sizes do not lead to statistically significant test
score gains. We estimate that reductions in class
size in kindergarten through second grade produce
a 6 to 11 percent annual real rate of return on
investment.
Finding: Full-Day vs. Half-Day Kindergarten
We analyze 23 rigorous evaluations and find that
full-day kindergarten, compared with half-day
kindergarten, produces a statistically significant
boost to test scores during, or shortly after,
kindergarten. These positive early gains, however,
appear to erode almost completely during grades
one through three. Thus, for full-day kindergarten
to generate long-term academic benefits, public
policies need to examine how to sustain the early
gains from any investments in full-day kindergarten.
Experimentation seems warranted.
These questions are, of course, difficult to answer
because in the real world few things can be known
with certainty. Determining causality, however, is
not a problem unique to education policy. Almost
all business and public policy decisions involve
different degrees of risk and uncertainty in knowing
whether desired outcomes can be secured with a
given strategy. In this report, we describe the
steps we have taken to identify whether the costs
of certain evidence-based K–12 policies and
programs are likely to relate to student outcomes.
Our analytical work is not yet complete; rather, this
is our first report describing progress to date.
Comments are welcomed.
For this current assignment on K–12 topics, the
Institute is building on its previous analyses of the
costs and benefits of other public policies. The
Washington State legislature has, in recent years,
directed the Institute to examine evidence-based
programs related to prevention, early intervention,
mental health, substance abuse treatment, and
criminal justice policies for both juveniles and
adults.3 In these previous studies, the legislature
also asked the Institute to estimate the costs and
benefits of research-based approaches.
This report begins by describing, briefly, our
research approach. We then present and discuss
our findings for the two K–12 topics covered in
this report: class size reductions and full-day
kindergarten. For readers interested in technical
matters, we also include an appendix, beginning
on page 16, that provides greater detail on the
Institute’s analytic procedures, economic
methods, and results.
Research Approach
In this initial review of K–12 topics, we focus on a
single type of educational outcome: student
academic performance. In addition to academic
skills, of course, public expectations place many
other goals on the K–12 system. These additional
goals include improved non-cognitive outcomes
such as promoting individual discipline and a work
ethic, citizenship, reduced criminal activity, reduced
drug and alcohol abuse, reduced teen pregnancy,
and so on.4 While these goals are important, our
initial review focuses on a narrower question: What
works to improve academic outcomes? This
outcome is especially timely, because state and
federal polices have placed student academic
performance as the prime outcome measure for the
K–12 system.
2
The types of academic outcomes that we analyze
depend on the specific measures used in the
existing K–12 evaluation studies we review. These
academic outcome measures include, but are not
limited to, the following:
9 Standardized test scores;
9 Course grades or grade point averages;
9 Grade retention;
9 Years in special education;
9 High school graduation/dropping out; and
9 Longer-range outcomes such as college
attendance, college graduation, employment,
and earnings.
Our research approach involves two general steps.
Step One: What Works? What Doesn’t? In order
to estimate whether a particular type of K–12
program or policy is likely to affect student
academic performance, we systematically assess
the findings of all methodologically sound research
studies we can locate. For each high-quality
evaluation we find, we compute an “effect size”—a
Legislative Study Direction
The 2006 Washington State Legislature directed
the Institute to initiate research that will provide
Washington with an on-going analysis of evidence-
based K–12 programs and services, as well as
cost-benefit analyses of each approach. The
language initiating the study was in Engrossed
Substitute Senate Bill 6386 §607 (15) which
directed the Institute to:
“…begin the development of a repository of
research and evaluations of the cost-benefits of
various K–12 educational programs and services.
The goal for the effort is to provide policymakers
with additional information to aid in decision making.
Further, the legislative intent for this effort is not to
duplicate current studies, research, and evaluations
but rather to augment those activities on an on-
going basis. Therefore, to the extent appropriate,
the institute shall utilize and incorporate information
from the Washington learns study, the joint
legislative audit and review committee, and other
entities currently reviewing certain aspects of K–12
finance and programs. The institute shall provide
the following: (a) By September 1, 2006, a detailed
implementation plan for this project; (b) by March 1,
2007, a report with preliminary findings; and (c)
annual updates each year thereafter.”
statistical summary measure indicating the degree
to which an evaluated policy or program changes
an academic outcome. Then, for a group of studies
on a particular K–12 topic, we combine the effect
sizes to determine whether, on average, outcomes
can be expected to change with the program or
policy under consideration.5
While it may be tempting to examine only one or
two studies on a topic, we think a restricted review
of existing research may lead to unrealistic or
biased expectations. By considering all
methodologically sound studies on a topic, our
approach seeks to determine the average
evidence-based effectiveness of each K–12 topic.
One always hopes for above-average
performance—a so called “Lake Wobegon” effect—
but for the K–12 taxpayer investments considered
in this review, we think it is more prudent to base
expectations on the average evidence-based result.
An analogy may help explain our approach: investing
in the stock market. If one is interested in knowing
the likely return from investing in the stock market, it
is better to examine the historical and expected
returns of many stocks rather than focusing on one
stock that has performed exceptionally well. Thus, a
broad stock market index like the S&P 500 provides
a more realistic gauge of expected stock market
returns than the historical return of any one
exceptional stock, such as Microsoft. One always
hopes for a Microsoft-like return, but expectations
are more likely to be fulfilled by anticipating the
average performance of many stocks.
Following this logic, for example, if one wants to
know whether a typical real-world investment in
preschool improves the academic outcomes for
low-income children, it is more prudent to assess
the results of all methodologically sound studies
that have been done on preschools for this
population (the equivalent of the S&P 500
approach) rather than selecting one preschool
study that happened to achieve exceptional returns
(the Microsoft analogy). Unless one has inside
knowledge of how to pick consistently the next
Microsoft, or confidence that schools can duplicate
regularly the all-time best preschool approach, then
it is safer to assume an average return based on a
larger group of results.
Thus, our approach to determining “What Works?”
is to review all of the methodologically sound
studies on a topic in order to estimate the likely
return on investment for a typical, real-world, K–12
program or policy.
We include studies in our review after screening for
methodological rigor and relevance for Washington
State. We include random assignment studies,
although there are relatively few of these “gold-
standard” studies. Therefore, we also include
rigorous quasi-experimental or observational
studies when special methodological care has been
taken to isolate the causal effect of a K–12 policy or
program on academic outcomes.
In the education field, paying close attention to a
study’s methodological quality appears to be
especially important because parents, students,
schools, and voters each exert a considerable
influence on how students and educational
resources are distributed. This real-world non-
random sorting of students and resources can make
it difficult for a study to isolate the causal effect of a
program or policy on student outcomes. A study
with very good data can statistically control for some
or perhaps many of these factors, but usually there
are other factors—unobserved to the researcher—
that can confound the ability of a study to identify
causal effects. Fortunately, as we discuss, there
have been recent advances in datasets, as well as
increased use of advanced statistical methods, that
have allowed researchers to improve their ability to
identify important outcomes of certain education
policies and programs.
Step Two: What Are the Expected Returns on
Investment? One of the precepts of economics is
that “there is no such thing as a free lunch.” Each
of the programs and policies discussed in this
report can cost taxpayers money. Therefore, in
addition to estimating whether research indicates
something works, it is also important to estimate
whether the benefits of an approach outweigh its
costs. In this study, we conduct an economic
analysis by stacking the expected monetary value
of any statistically significant benefits against the
costs of the program or policy. To do this, we have
developed, and are continuing to refine, techniques
to measure costs and benefits associated with the
outcomes of K–12 programs, policies, and services.
We use the findings from recent economic research
to provide a range of estimates of the benefits of
statistically significant educational outcomes. We
model these outcomes in a “human capital”
framework. Economists such as Alan Krueger and
Eric Hanushek, who often disagree on whether
certain K–12 policies achieve outcomes, generally
use a similar human capital approach to monetize
the benefits of any outcomes obtained.6 In the
3
human capital model, successful investments in
K–12 policies and programs (i.e., investments that
have an evidence-based ability to boost academic
performance), are estimated to generate benefits
over a number of years into the future. The
benefits typically include labor market and other
types of non-market benefits. We summarize these
monetary costs and benefits with the usual set of
financial summary statistics: net present values,
benefit-to-cost ratios, and rates of return on
investment.
As in our previous cost-benefit analyses, we
estimate life-cycle costs and benefits from two
perspectives: first, we estimate the benefits that
accrue directly to program participants (in this case,
the students), and second, we estimate the benefits
that accrue to non-participants.
For example, a student who scores higher on
standardized tests can be expected to enjoy the
benefit of greater earnings in the labor market
compared with students who do not score as well.7
Non-participants benefit from the taxes paid on
those increased earnings. Economists have also
been examining whether improved K–12 outcomes
are related to other desirable outcomes such as:
reduced crime; improved health care and lower
health care costs; reduced foster care; so-called
“knowledge spillovers” that stimulate general
economic growth; and increased civic participation.8
While the research underlying many of these non-
market outcomes is more uncertain and less well
developed than the labor market outcomes, we
conduct sensitivity analyses to test how the range
of total benefits might be affected by successful
K–12 educational policies.
The appendix to this report describes our economic
procedures in detail.
4
K–12 Topics Scheduled for Evidence-Based
Reviews and Cost-Benefit Analyses
The assignment from the Legislature was “to
begin the development of a repository of
research and evaluations of the cost-benefits of
various K–12 programs and services.” In
addition to the two topics covered in this initial
report, we have also begun work on a number of
additional topics by collecting and analyzing the
relevant research literature. The work on several
of these additional topics is underway but not yet
complete; results will be presented in upcoming
reports.
Topics for upcoming reports include, but are not
limited to the following:
9 Preschool education
9 Dropout prevention programs
9 Professional development activities
9 Effect of school size
9 Teacher quality effects
9 Alternative education programs
9 English language learner programs
9 Charter schools
9 Vouchers
9 Other early learning approaches
9 Teacher aides
9 Mentors for students
9 Mentors for new teachers
9 Tutoring
9 Teacher compensation
9 Summer school
9 Extended day/weekend programs
9 Grade retention
 Reductions in K–12 Class Size
Research Questions. Does reducing the
number of students in a classroom improve
student academic performance? If so, then by
how much? Are class size reductions more
effective in the lower grades or in middle school or
high school? Do students from lower income
families benefit more by class size reductions
than students from higher income families?
In addition to these questions of effectiveness,
there are also economic questions. Since it can
cost over $200 per student per year to reduce
class size by one unit, and since there are about
one million students in Washington’s public K–12
system, a system-wide reduction in class size by
just one unit could cost taxpayers about $200
million per year. This would represent about a 2.5
percent increase in statewide K–12 expenditures.
Thus, a significant economic question asks
whether there is solid empirical evidence that any
benefits of class size reductions would exceed
costs. Moreover, are there approaches other than
reducing class sizes that would produce a bigger
bang for the buck (where “bang” is measured as
gains in student academic performance)?
Background. Many of these class size questions
have been studied throughout the United States
and abroad since the 1960s. Despite this long
trail of research, however, the answers that have
been suggested remain controversial to many of
the researchers involved in the debate.9 As a
result, the class size issue continues to be an
active area of inquiry. The debate remains
pertinent, because proposals to reduce class
sizes as a means to improve student outcomes
are often put forward and adopted.10
Part of the controversy on this topic stems from
the nature of the early studies conducted on the
effects of class size. Many of the early studies
were based on simple relationships between class
size and student outcomes. As discussed earlier,
particularly in the education field, correlation may
not indicate causation. Parents, students,
schools, and voters exert a considerable influence
on how students and educational resources are
distributed in the K–12 system. This non-random
sorting of students and resources can make it
difficult for a correlation-based study to isolate the
true causal effect of reducing class sizes on
student outcomes. Even a study that statistically
controls for many factors often cannot adjust for
other telling factors that are unobserved to the
researcher, unless special statistical procedures
are employed. Most of the early studies suffered
from these sorts of statistical problems.
Therefore, it is perhaps not surprising that
interpreting the results of these early studies has
engendered controversy.11
Some of these methodological concerns could be
overcome with well-designed random assignment
studies. Unfortunately, random assignment studies
are infrequently used in the education field because
they are often expensive, difficult to conduct, and
they can raise ethical questions in deciding who
gets an intervention and who does not. There has
been, however, one important random assignment
study in the education field—the well known
Tennessee STAR experiment in reducing class
size. This experimental study is widely cited and
provides valuable lessons.12 Even this “gold-
standard” study, however, has been criticized for
not being a perfectly implemented random
assignment experiment and for the difficulty in
generalizing the results to the conditions of
everyday classrooms.13
Fortunately, in the last decade, there have been a
number of quasi-experimental studies estimating
the effect of class size reductions that have used
significantly improved statistical methods. Some
recent studies have also used new and improved
state, national, and international datasets. These
more recent studies represent substantial
improvements over the earlier correlation-based
studies. In our review of the research on class size,
we include both the results of the Tennessee STAR
experiment and the recent high-quality quasi-
experimental research studies. We think that,
combined, this group of studies forms the best
research evidence to date from which to draw
cause-and-effect conclusions about the effect of
reductions in class size on student academic
performance.
Literature Search. In conducting a review of the
research, the first task is to locate the relevant
studies. We began our search for evaluations of
the effects of class size by reading the citations in
studies known to us. This was followed by
5
searching the internet and academic library unit change in class size for the grade level in
information systems for published or yet-to-be which the resources were spent.
published studies. We then read and screened all
prospective studies for methodological rigor and About 49 percent of the separate tests were from
relevance for Washington State policy questions. studies conducted in the United States and the
Individual authors of the studies frequently needed remaining were of populations outside the United
to be contacted to obtain additional information. States. We excluded international studies where
We found 38 class size studies with sufficient class sizes are at substantially different levels than
methodological rigor to include in our analysis. The those found in the United States. As we describe,
citations to these studies are listed in Exhibit 5. we also tested to determine whether results from
United States class size studies are significantly
Characteristics of the Studies Included in Our different from international studies.
Review. Exhibit 1 lists some of the characteristics
of the 38 methodologically sound studies included In terms of methodology, about 62 percent of these
in our review. The majority of the studies were effects were from studies that employed an
written or published recently. The oldest study was instrumental variables or regression discontinuity
published in 1989, seven were published between design; 20 percent used a correlation-based design
1995 and 1999, and 30 were published from 2000 (a hierarchical linear model or ordinary least
to 2006. squares) with rich datasets that allowed the
researchers to include a considerable number of
Exhibit 1 statistical controls; about 6 percent used a fixed
Description of Studies Included effects panel data approach without an instrumental
Number Pct. variable; and there were two random assignment
Number of studies included in analysis 38 100% studies. For the Tennessee STAR study, we
Publication Date included two reports that independently analyzed
1980s 1 3% the data from this important study.14
1995 to 1999 7 18%
2000 to 2006 30 79%
Of these 69 separate tests, 97 percent directly
Number of grade-level effect sizes (ES) 69
measured whether standardized test scores were
Domestic or International grade-level ES influenced by changes in class size, and about 3
United States 34 49%
percent measured whether high school graduation
International 35 51%
rates were influenced by class size changes. We
Methodology of grade-level ES also examined studies testing whether changes in
Instrumental variables (IV) or
regression discontinuity design
43 62% K–12 spending influenced standardized test scores;
Hierarchical linear model or ordinary seven of the 69 separate tests (10 percent) in our
14 20%
least squares regression analyses were of this form. Even though this last
Fixed effects regression without an IV 4 6% group of studies does not measure class size
Random assignment 8 12%
directly, we included their findings because a high
Outcome variable: student test scores 67 97% proportion of K–12 operational spending is for
Outcome variable: other (graduation) 2 3% teaching staff and, therefore, expenditures are
Policy variable: class size change 62 90%
Policy variable: K–12 spending 7 10% probably a reasonable proxy for changes in class
size. We did, however, conduct our overall analysis
Grade when the resources were spent to with and without this last group of studies included.
lower class size
Kindergarten through third grade 18 26%
Fourth through sixth grade 25 36%
Twenty-six percent of the 69 tests were for class
Seventh through eighth grade 18 26% size reductions primarily in kindergarten through
Ninth through twelfth grade 8 12% second grade. Thirty-six percent were for
reductions in grades three through six, 26 percent
occurred during seventh through eighth grades, and
These studies contributed 69 separate tests of 12 percent during high school.
whether reductions in K–12 class sizes affect
student academic performance. The reason there
are more tests than studies is that some studies
estimated results of class size reductions for
different grade levels. In our analysis, the unit of
observation is the estimated effect size for a one-

6
 Exhibit 2
Changes in Academic Achievment
From Reducing Class Size by One Unit
.05

Gains on Average Math and Reading Tests

.04

.03

Effect Size .02

.01

.00

-.01 Each circle is

-.02 the effect
size from a
-.03 study, N = 69
-.04
0K 1 2 3 4 5 6 7 8 9 10 11 12
Grade in Which the Class Size Was Reduced

Results and Findings. To measure results, we second grade is consistently associated with
calculate an “effect size” for each of the 69 positive gains in academic test scores. For third
separate tests contributed by the 38 studies in our through sixth grade, the results are more mixed
review. An effect size is a statistical summary with some studies indicating positive results and
measure describing the degree to which academic some indicating lower or negative results. By
performance is improved as a result of a reduction middle school and high school, the effects appear
in class size. The bigger the effect size, the bigger to be small, on average, and there have been some
the impact. An effect size of zero means there is studies indicating no gain or even a reduced level
no effect of the class size reduction on test scores. of academic performance with reduced class sizes.
For technical readers, the appendix describes the It is also clear that in middle school, the raw results
procedures we use to calculate effect sizes. are quite varied, while in high school there are
relatively few rigorous studies that have tested the
An effect size measures the expected change in effect of class size reductions.
test scores, expressed in standard deviation units.
Washington’s standardized test is the Washington Thus, the simple plot of effect sizes in Exhibit 2
Assessment of Student Learning (WASL). The reveals that class size reductions in the early
average student-level score on the 2006 10th-grade grades are likely to be more effective than during
math WASL was 401 with a standard deviation of higher grades.
38. Thus, for example, an educational policy that
produces a large effect size of 0.5 would mean an We then examine these 69 raw effect sizes with
average gain of 19 points on the WASL (19 = .5 X multivariate regression. The purpose of this more
38), or about a 4.7 percent change in average test in-depth analysis is to refine the simple plot shown
scores (.047 = 19 / 401). in Exhibit 2 by controlling for the characteristics of
the studies. As shown in Exhibit 1, some of the
In our analysis of class size, we calculate the effect studies were from the United States, some were
of a one-unit change in class size on test scores. from international locations; some used certain
For example, our effect sizes measure the change types of statistical identification methodologies,
in standard deviation units by moving from a class others did not; some used student-level data,
size of 20 to a class size of 19.15 others used class- or district-level data. Using
standard statistical procedures, we also weight the
Exhibit 2 displays a simple plot of the 69 effect results of the different studies so that a study that
sizes arranged by the grade in which the class size evaluated many students is given more weight than
reduction took place. Each dot represents an effect a study that evaluated far fewer students. Our
size from an individual study and measures the multivariate analyses allow us to test for the
change on average math and/or reading tests. A significance of these factors. In the appendix we
simple examination of Exhibit 2 indicates that describe our methods and results in technical
reducing class sizes in kindergarten through the detail.

7
 Exhibit 3
Effect of Class Size Reductions
.030
T he re d bo xe s a re t he
Effect size for a one-unit
.025
e s t im a t e d e f f e c t
drop in class size
.020 s ize s ; t he v e rt ic a l
.019
line s a re 9 5 %
.015 c o nf ide nc e int e rv a ls .
.010
.007
.005
.000
-.001
-.005
-.010
K to 2 3 to 6 7 to 8
Grade Level When Class Size Reduced
Our estimated effects from our preferred regression
model are presented in Exhibit 3. The findings are
consistent with those in Exhibit 2. There are
statistically significant effects for two grade levels:
kindergarten through second grade, and third
through sixth grade, although the effects for the
latter group are just 35 percent of the effects for the
kindergarten through second grade group. The
results for middle and high schools indicate that
class size reductions do not generate statistically
significant improvements in test scores—note that
the 95 percent confidence intervals shown in
Exhibit 3 for these two grade level groups include
zero as a possibility.
Return on Investment (ROI) Calculations. The
purpose of this study is to estimate the costs and
benefits of K–12 policies and programs. We
calculate a return on investment statistic that is
computed in the same general way as that for
private sector investments.
In the appendix, we describe in detail the
procedures we use to estimate the monetary
benefits associated with the effect sizes we just
discussed. We estimate that increased test scores
generate monetary benefits beginning at age 18
when the student would begin to be attached to the
labor market. We provide a range of returns on
investment, since there are several factors that can
be estimated only with uncertainty. In particular, we
varied these factors (details shown in Exhibit B.2 in
the appendix):
1) The estimate of the initial gain is test score
effect size, shown in Exhibit 3;
2) An annual rate of decay in this effect size to
the end of high school;
3) An average annual real rate of growth in labor
market earnings;
8
4) An estimate of the effect of a gain in test
scores on lifetime earnings in the labor market;
5) Alternative social rates of return to account for
such non-labor market factors as reduced
crime, reduced health care costs, increased
civic participation, and “knowledge spillovers”
that stimulate general economic growth;
6) Alternative real discount rates.
ROI Finding: Class Size Reductions in
.004
Kindergarten Through Grade Two. As shown in
Exhibit 3, kindergarten through grade two are the
grade levels for which we estimate the largest
effects on test scores. We estimate that a one-unit
9 to 12 drop in class size for these grades would cost about
$217 per student per year to pay for the increased
operating and capital costs. We estimate that the
real internal rate of return on investment for a one-
unit drop in class size during kindergarten through
second grade ranges from 5.7 to 11 percent. The
average return on investment is 8.3 percent.16
Expressed in terms of an average benefit-to-cost
ratio, this investment generates $2.79 in benefits for
each dollar of cost.
For comparison purposes, the long-run annual rate
of return on investment for the equities that make
up the S&P 500 stock market index is about 4.4
percent per year.17
ROI Finding: Class Size Reductions in Grades
Three Through Six. Exhibit 3 indicates that class
size reductions in grades three through six
generate a significant—but lower—effect on test
scores than in the first few grades. This reduced
effect means lower returns on investment. We
estimate that the real average return on investment
for a one-unit drop in class size during grades three
through six is about 6 percent, or $1.38 in benefits
per dollar of cost.
ROI Finding: Class Size Reductions in Middle
School and High School. As shown in Exhibit 3,
we did not find statistically significant effects for
class size reductions in middle and high school, so
we did not compute return on investment estimates.
Additional Analysis of Low-Income Populations.
Some of the studies in our review include
information on whether students from low-income
families fare better with class size reductions than
students from non-low-income families. We
conducted an additional analysis for this group of
studies, and Exhibit 4 plots effect sizes against the
percentage of low-income students reported in
these studies. The effect of class size reductions
appears greater in classes with larger proportions
of students from low-income families. Our
statistical analysis of this relationship reveals that
students from low-income families benefit more
from class size reductions than students from
higher-income families. That is, the effect size is
larger for studies where a higher percentage of
students were from low-income families. Appendix
C.1 provides details of this analysis.

Exhibit 4
Class-Size Reductions by Income Level
.045
Effect size for a one-unit

.040
.035
drop in class size

.030
.025
.020
.015
.010
.005
.000
-.005
0% 20% 40% 60% 80%

Percent Students From

Low-Income Families

9
 Exhibit 5
Citations to the Studies Used in the Statistical Analyses of Class Size Reductions
(Some studies contributed independent effect sizes from more than one location or grade level)

Akerhielm, K. (1995). Does class size matter? Economics of Education Review, 14(3): 229-241.
Angrist, J. & Lavy, V. (1999). Using Maimonides' Rule to estimate the effect of class size on children's academic achievement. Quarterly Journal of
Economics, 114(2): 533-576.
Blatchford, P., Goldstein, H., Martin, C., & Browne, W. (2002). A study of class size effects in English school reception year classes. British Education
Research Journal, 28(2): 169-185.
Bonesrønning, H. (2003). Class size effects on student achievement in Norway: Patterns and explanations. Southern Economic Journal, 69(4): 952-965.
Borland, M.V., Howsen, R.M., & Trawick, M.W. (2005). An investigation of the effect of class size on student achievement. Education Economics, 13(1):
73-83.
Bressoux, P., Kramarz, F., & Prost, C. (2005). Teachers' training, class size and students' outcomes: Evidence from their grade classes in France. Paris,
France: Center for Research in Economics and Statistics.
Browning, M. & Heinesen, E. (2005). Class size, teacher hours and educational attainment (CAM Working Papers No. 2003-1). Copenhagen, Denmark:
University of Copenhagen, Department of Economics, Centre for Applied Microeconometrics.
Dearden, L., Ferri, J., & Meghir, C. (2002). The effect of school quality on educational attainment and wages. Review of Economics and Statistics, 84(1):
1-20.
Dustmann, C., Rajah, N., & van Soest, A. (2003). Class size, education, and wages. The Economic Journal, 113(485): F99-F120.
Ecalle, J., Magnan, A., & Gibert, F. (2006). Class size effects on literacy skills and literacy interest in first grade: A large-scale investigation. Journal of
School Psychology, 44(3): 191-209.
Feinstein, L. & Symons, J. (1999). Attainment in secondary school. Oxford Economic Papers, 51(2): 300-321.
Fergeson, R.F. & Ladd, H.F. (1996). How and why money matters: An analysis of Alabama schools. In H.F. Ladd (Ed.), Holding schools accountable:
Performance-based reform in education (pp. 265-298). Washington: Brookings Institution.
Fuchs, T. & Wößmann, L. (2004). What accounts for international difference in student performance? A re-examination using PISA data (Report no.
1287). Bonn, Germany: Institute for the Study of Labor (IZA).
Grissmer, D.W. & Flanagan, A. (2006). Improving the achievement of Tennessee students: Analysis of the National Assessment of Educational Progress.
Santa Monica, CA: RAND.
Grissmer, D.W., Flanagan, A., Kawata, J., & Williamson, S. (2000). Improving student achievement: What state NAEP test scores tell us. Santa Monica,
CA: RAND.
Guryan, J. (2003). Does money matter? Estimates from education finance reform in Massachusetts. Chicago, IL: University of Chicago, Graduate School
of Business.
Haegeland, T., Raaum, O., & Salvanes, K.G. (2005). Pupil achievement, school resources and family background (Report No. 1459). Bonn, Germany:
Institute for the Study of Labor (IZA).
Hoxby, C.M. (2000). The effects of class size on student achievement: New evidence from population variation. The Quarterly Journal of Economics,
115(4): 1239-1285.
Iacovou, M. (2002). Class size in the early years: Is smaller really better? Education Economics, 10(3): 261-290.
Jakubowski, M. & Sakowsi, P. (2005). Quasi-experimental estimates of class size effects in primary school in Poland. Warsaw, Poland: Warsaw
University, Faculty of Economics.
Jenkins, A., Levacic, R., & Vignoles, A. (2006). Estimating the relationship between school resources and pupil attainment at GCSE (Report No. RR727).
London: University of London, Institute of Education.
Jepsen, C. & Rivkin, S. (2002). Class size reduction, teacher quality, and academic achievement in California public elementary schools. San Francisco,
CA: Public Policy Institute of California.
Kang, C. (2005). Effects of small classes on academic achievement: Evidence from new entrants to Project STAR. Singapore: National University of
Singapore, Department of Economics.
Kinnucan, H.W., Zheng, Y., & Brehmer, G. (2006). State aid and student performance: A supply-demand analysis. Education Economics, 14(4): 487-509.
Krueger, A. (1999). Experimental estimates of education production functions. Quarterly Journal of Economics, 114(2): 497-532.
Levacic, R., Jenkins, A., Vignoles, A., Steele, F., & Allen, R. (2005). Estimating the relationship between school resources and pupil attainment at Key
Stage 3 (Report No. RR679). London: University of London, Institute of Education.
Levin, J. (2001). For whom the reductions count: A quantile regression analysis of class size and peer effects on scholastic achievement. Empirical
Economics, 26(1): 221-246.
McGiverin, J., Gilman, D., & Tillitski, C. (1989). A meta-analysis of the relation between class size and achievement. The Elementary School Journal,
90(1): 47-56.
Molnar, A., Smith, P., Zahorik, J., Palmer, A., Halbach, A., & Ehrle, K. (1999). Evaluating the SAGE program: A pilot program in targeted pupil-teacher
reduction in Wisconsin. Educational Evaluation and Policy Analysis, 21(2): 165-177.
NICHD Early Child Care Research Network. (2004). Does class size in first grade relate to children's academic and social performance or observed
classroom processes? Developmental Psychology, 40(5): 651-664.
Papke, L.E. (2006). The effects of changes in Michigan's school finance system. East Lansing, MI: Michigan State University, Department of Economics.
Pikkety, T. (2004). Should we reduce class size or school segregation? Theory and evidence from France. Paris, France: ENS-EHESS (as described in
Valdenaire, 2004).
Ready, D.D. & Lee, V.E. (2006). Optimal context size in elementary schools: Disentangling the effects of class size and school size. Washington, DC:
Brookings Papers on Education Policy.
Rivkin, S.G., Hanushek, E.A., & Kain, J.F. (2005). Teachers, schools, and academic achievement. Econometrica, 73(2): 417.
Sander. W. (1999). Endogenous expenditures and student achievement. Economics Letters, 64(2): 223-231.
Urquiola, M. (2006). Identifying class size effects in developing countries: Evidence from rural Bolivia. The Review of Economics and Statistics, 88(1): 171-
177.
Valdenaire, M. (2006). Do younger pupils need smaller classes? Evidence from France. London: London School of Economics, Centre for Economic
Performance.
Wößmann, L. & West, M.R. (2006). Class size effects in school systems around the world: Evidence from between-grade variation in TIMSS. European
Economic Review, 50(3): 695-736.

10
 Full-Day vs. Half-Day Kindergarten
Research Questions. Do children who attend
full-day kindergarten exhibit greater academic
gains than children who attend half-day
kindergarten? If so, how big are the gains? Is
full-day kindergarten of greater benefit for minority
and low-income students? Are the gains
sustained as children progress in school?
We also ask economic questions. We estimate
that full-day kindergarten costs about $2,611
more per child than half-day programs to pay for
changes in operating and capital costs. Providing
full-day kindergarten to all children in Washington
could increase state education expenditures by
$190 million. Are the academic benefits of full-
day kindergarten worth the additional cost of
these programs?
Background. When kindergartens were first
introduced in the United States, they were full-day
programs. Later, during the Second World War
when there was a teacher shortage, kindergarten
programs were shortened to half-days and
children attended either morning or afternoon
programs.18 In the 1960s, more schools began to
implement full-day programs, particularly to
enhance the school-readiness of disadvantaged
children. The trend toward full-day programs has
continued. Nationally, between 1970 and 2000,
the percentage of kindergartners attending full-
day programs increased from 14 percent19 to over
60 percent. 20
Across the United States, decisions about offering
full- or half-day kindergarten are made primarily at
the local level. In 2005, nine states required local
districts to offer full-day kindergarten.21
In Washington State, half-day kindergarten is
funded by the state general appropriation. Some
districts have chosen to offer full-day programs
funded by local levies, parent fees, or other non-
designated sources. During the 2006-07 school
year, 37 percent of kindergartners in Washington
public schools attended full-day programs.22
The relative merits of full-day kindergarten—
compared with half-day kindergarten—have been
studied in the United States since the 1960s.
Despite this long history of research, however,
there is still controversy about the long-term
academic benefits of full-day programs. The
majority of studies have focused on the academic
gains of children at the end of kindergarten. More
recent studies, however, have included longer-
term follow-up periods enabling researchers to
examine whether academic gains persist in the
early years of education. For example, the Early
Childhood Longitudinal Study—Kindergarten
Class of 1998–99 (ECLS-K) is a nationally
representative study following a sample of 20,000
children enrolled in kindergarten in 1998. 23
These public-use data have allowed three
independent evaluations of the longer-term effects
of full-day vs. half-day kindergarten.
Literature Search. In conducting a review of the
research, the first task is to locate the relevant
studies. We began our search for evaluations of
the effects of full-day kindergarten by reading the
citations in studies known to us. This was followed
by searching the internet and academic library
information systems for published or yet-to-be
published studies. We then read and screened all
prospective studies for methodological rigor and
relevance to Washington State policy questions.
Individual authors of the studies frequently were
contacted to obtain additional information. We
found 23 studies with sufficient methodological rigor
to include in our analysis. The citations to these
studies are listed in Exhibit 10.
We only include studies of full-day kindergarten that
have a comparison group of children who attended
half-day kindergarten. We would expect children to
show an increase in learning over the course of the
kindergarten year; our research question is to find
out whether a full-day program enhances the
learning we would expect from a half-day program.
We exclude some studies with comparison groups
if the authors did not make clear how children were
chosen for the full- and half-day programs,
particularly if the analysis did not control for
demographics or children’s skills at the start of
kindergarten. These controls are especially
important if full-day programming is optional,
because parents who opt for full-day kindergarten
may be different in ways that influence their
children’s academic performance from parents who
choose half-day kindergarten.
11
Full-day kindergarten is often one part of a
remediation package for children at risk of
academic difficulty. In addition to full-day
kindergarten, some interventions included class
size reductions, bilingual instruction, or additional
classroom aides. In most cases, we excluded
these studies, because we could not isolate the
effects of full-day kindergarten. We did include one
study where the only other intervention was a
reduction in class size. In that case we adjusted
results using our effect size for smaller classrooms
described in the previous section of this report.
Characteristics of the Studies Included in Our
Review. Exhibit 6 lists some characteristics of the
23 methodologically sound studies included in our
review. As we saw with the class size literature, the
majority of studies are recent. Six were published
before 1990, including the oldest study published in
1970. Of the 11 studies published after 2000, five
are independent analyses of the ECLS-K survey
data; between them they provide follow-up
information through the fifth grade. All of the studies
were conducted in the United States. Citations for
these studies are provided in Exhibit 10.
Exhibit 6
Description of Studies Included
Number Pct.
Number of studies included in analysis 23 100%
Publication Date
1970 to 1990 6 26%
1991 to 2000 6 26%
2001 to 2006 11 48%
32
Number of grade-level effect sizes (ES)
Domestic or International grade-level ES
United States 32 100%
International 0 0%
Grade when the effects were measured
Kindergarten 17 53%
First grade 6 19%
Second grade 2 6%
Third grade 3 9%
Fourth grade 3 9%
Fifth grade 1 3%
The 23 studies provide results for 22 distinct
groups of children. The number of studies and
number of groups are not the same because the
five ECLS-K studies report on the same
population of children at different times after
kindergarten, and some studies report on more
than one group. Altogether, the studies provide
32 effect sizes for use in our analysis. We have
more effect sizes than studies because some of
the studies measured outcomes at several times
after the end of kindergarten.
12
All of the studies reported results on standardized
tests. While some reported other results, such as
attendance, behavior, and teacher and parent
satisfaction, our analysis includes just academic
achievement. Of course, these other outcomes
are important, but they are beyond the focus of
this initial report.
Many of the programs evaluated in the studies
were targeted at disadvantaged children in inner-
city schools. Some, but not all, reported the
percentage of poor or minority children in the
schools. Four studies reported on schools where
children were predominantly white, located in
upper-middle-class neighborhoods.
Results and Findings. To measure results, we
calculate an “effect size” for each of the 32
separate tests contributed by the 23 studies in our
review. An effect size is a statistical summary
measure describing the degree to which academic
performance is improved as a result of lengthening
the kindergarten program from half- to full-day.
The bigger the effect size, the bigger the impact
that full-day kindergarten is estimated to have on
standardized test scores. An effect size of zero
means there was no effect of full-day kindergarten
on test scores. For technical readers, Appendix A
describes the procedures we use to calculate
effect sizes.
An effect size measures the expected change, in
standard deviation units, in test scores. For
example, Washington’s standardized test is the
Washington Assessment of Student Learning
(WASL). The average student-level score on the
2006 4th grade math WASL was 406 with a
standard deviation of 37. Thus, for example, an
educational policy that produced a large effect size
of 0.5 would mean a gain of 18.5 points on the
WASL (18.5 = .5 X 37), or about a 4.6 percent
change in average test scores (.046 = 18.5 / 406).
Our results are consistent with the findings of others:
Full-day kindergarten provides a significant effect by
the end of kindergarten. This effect size, based on all
17 populations measuring effects during kindergarten,
is 0.181. The result is statistically significant because,
as shown in Exhibit 7, the 95 percent confidence
interval does not include zero. Using the analogy
above, and assuming no decay in effect over time,
this effect size would result in a 4th-grade math
WASL score of 413, or about a 1.7 percent increase
in average test scores.
 Research has shown that low-income and minority Because full-day kindergarten is often offered to
students are more often disadvantaged by the time disadvantaged children, we also analyzed from five
they begin school.24 Many school districts have studies reporting on low-income and minority children
employed full-day kindergarten programs to better separately. Exhibit 9 shows the results at kindergarten
prepare these groups for first grade. Several and later for these groups.27 The exhibit indicates that
studies evaluated the results for low-income and the short-term benefits decrease significantly, in a
minority students. By the end of kindergarten, the pattern similar to that of the entire sample.
effects for disadvantaged children were about the
same as for the entire sample (Exhibit 7).25
Exhibit 9
Effect Size Decay for
Exhibit 7 Disadvantaged Groups
0.25
Effect Size at End of Kindergarten
.300 Poverty
0.2
Black
.250
0.15
Hispanic
.200

Effect Size
Head Start
Effect Size

.181 0.1
.168
.150
0.05

T he re d bo xe s a re t he
.100
e s t im a t e d e f f e c t 0
s ize s ; t he v e rt ic a l
.050 line s a re 9 5 % -0.05
c o nf ide nc e int e rv a ls .
.000
-0.1
Entire Sam ple Disadvantaged
K 1 3
Children
Grade When Effect Was Measured

Based on our analysis, it is clear that full-day

Do these early gains persist? As noted, the kindergarten provides academic benefits by the
question of whether these early gains in full-day end of the kindergarten school year but that the
kindergarten are sustained is central to determining effects erode almost completely in grades one
the return on investment. While the results at the through three.
end of kindergarten are statistically significant,
Exhibit 8 indicates that the gains are no longer
evident by the end of first grade.26 There were 15
effects that measured academic success beyond
kindergarten. Exhibit 8 shows that these results are
insignificant because the 95 percent confidence
intervals include zero effect as a possibility.

Exhibit 8
Effect Size Decay
.300
T he re d bo xe s a re t he
.250 e s t im a t e d e f f e c t
.200 s ize s ; t he v e rt ic a l
.181 line s a re 9 5 %
Effect Size

.150 c o nf ide nc e int e rv a ls .

.100

.050 .048
.000 .011 .000

-.050
-.100
K 1 2 to 3 4 to 5
Grade When Effect Was Measured

13
Why do the benefits erode so quickly? An effect
size statistic measures the difference between
children in the two kindergarten schedules. Thus,
the decrease in effect size could be due to losses
by full-day children in the years after kindergarten,
greater gains made by half-day children, or some
combination of the two. DeCicca’s (2006) analysis
of ECLS-K data for Black, Hispanic and White
children found evidence of a “summer fallback”
between the end of kindergarten and the start of
first grade. This summer effect was especially
noticeable for Black children.
The lesson seems to be that for full-day kindergarten
to generate long-term academic benefits, public
policies need to examine how to sustain the early
gains from any investments in full-day kindergarten.
We estimate that extending kindergarten schedules
from half- to full-day costs $2,611 per student per
year to pay for the increased operating and capital
costs. If public policies can be found that sustain
the initial test score gains of full-day kindergarten
(shown in Exhibit 7), then we estimate the present
value of the benefits to be $5,600. These benefits
would represent the lifetime gains in earnings and
other benefits if the early test score gains could be
maintained. Of course, the programs necessary to
sustain the full-day kindergarten gains would not be
free, so from the $5,600 advantage one would need
to subtract the costs of these supplemental
programs, in addition to subtracting the $2,611 cost
of full-day kindergarten.

Return on Investment Calculations. The purpose

of this study is to estimate the costs and benefits of
K–12 policies and programs. Without sustained
benefits beyond the end of kindergarten, we would
estimate no long-term financial benefits for full-day
kindergarten. Thus, the net result is a negative
benefit of -$2,611—our estimated per-student cost
of full-day vs. half-day kindergarten.

If, on the other hand, public policies can be

implemented that sustain the early gains in test
scores of full-day kindergarten, then there would be
significant net long-term benefits.

To estimate the potential net benefits that could

be obtained if full-day kindergarten’s short-term
gains can be sustained to the end of high school,
we calculate a return on investment statistic. We
use the same economic model we describe in
our discussion of class size reduction and in
Appendix B.

We provide a range of returns on investment since

there are several factors that can be estimated only
with uncertainty. In particular, we varied these factors
(details shown on Exhibit B.2 in the appendix):

1) The estimated initial gain effect size from full-day

kindergarten, shown on Exhibit 7;
2) An average annual real rate of growth in labor
market earnings;
3) An estimate of the effect of a gain in test scores
on lifetime earnings in the labor market;
4) Alternative social rates of return to account for
such non-labor-market factors as reduced crime,
reduced health care costs, increased civic
participation, and “knowledge spillovers” that
stimulate general economic growth;
5) Alternative real discount rates.

14
 Exhibit 10
Citations to the Studies Used in the Statistical Analyses of Full-Day vs. Half-Day Kindergarten
(Some studies contributed independent effect sizes from more than one location or grade level)

Amsden, D., Buell, M., Paris, C., Bagdi, A., Cureval, T., Edwards, N., et al. (2005). Delaware pilot full-day kindergarten evaluation: A comparison of ten full-day
and eight part-day kindergarten programs, School year 2004-2005. Newark, DE: University of Delaware, Center for Disabilities Studies.
Cannon, J.S., Jacknowitz, A., & Painter, G. (2006). Is full better than half? Examining the longitudinal effects of full-day kindergarten. Journal of Policy Analysis
and Management, 25(2): 299-321.
Carapella, R. & Loveridge, R.L. (1978). A comparative report of the achievement of the kindergarten extended day program. St. Louis, MO: St. Louis Public
Schools.
DeCicca, P. (2007). Does full-day kindergarten matter? Evidence from the first two years of schooling. Economics of Education Review, 26(1): 67-82.
Del Gaudio Weiss, A. M., & Offenberg, R. M. (n.d.) Differential impact of three types of kindergarten experience on students' academic achievement through
third grade. Philadelphia, PA: School District of Philadelphia, Office of Research & Evaluation.
Del Gaudio Weiss, A.M., & Offenberg, R.M. (n.d.). Differential impact of type of kindergarten experience on academic achievement and cost-benefit through
grade 4: An examination of four cohorts in a large urban school district. Philadelphia, PA: School District of Philadelphia, Office of Research & Evaluation.
Elicker, J. & Mathur, S. (1997). What do they do all day? Comprehensive evaluation of full-day kindergarten. Early Childhood Research Quarterly, 12: 459-80.
Entwisle, D., Alexander, K.L., Cadigan, D., & Pallas, A.M. (1987). Kindergarten experience: Cognitive effects or socialization? American Educational Research
Journal, 24(Autumn): 337-364.
Evans, E.D. & Marken, D. (1983). Longitudinal follow-up comparison of conventional and extended-day public school kindergarten programs. Paper presented
at the annual meeting of the American Educational Research Association, New Orleans, April (ERIC No. ED254298).
Hildebrand, C. (1997). Effects of all-day and half-day kindergarten programming on reading, writing, math, and classroom social behaviors. National Forum of
Applied Educational Research Journal, 10E(3): 14.
Holmes, C.T. & McConnell, B.M. (1990). Full-day versus half-day kindergarten: An experimental study. Paper presented at the annual meeting of the American
Educational Research Association, Boston, April (ERIC No. ED369540).
Le, V., Kirby, S.N., Barney, H., Setodji, C.M., & Gerswhin, D. (2006). School readiness, full-day kindergarten, and student achievement: An empirical
investigation. Santa Monica, CA: RAND Corporation.
Lee, V.E., Burkam, D.T., Ready, D., Honigman, J., & Meisels, S.J. (2006). Full-day versus half-day Kindergarten: In which program do children learn more?
American Journal of Education, 112(2): 163-208.
Morrow, L.M., Strickland, D.S., & Woo, D.G. (1998). Literacy instruction in half- and whole-day kindergarten. Newark, NJ: International Reading Association.
Nielsen, J., Cooper-Martin, E. (2002). Evaluation of the Montgomery County public schools assessment program: Kindergarten and grade 1 reading report.
Rockville, MD: Montgomery County Public Schools, Office of Shared Accountability.
Nunnelley, J. (1996). The impact of half-day versus full-day kindergarten programs on student outcomes: A pilot project (ERIC No. ED396857).
Park Hill School District. (1998). Full-day kindergarten 1997-98 evaluation report. Kansas City, MO: Park Hill School District, Office of Research, Evaluation,
and Assessment.
Saam, J., Nowak, J.A. (2005). The effects of full-day versus half-day kindergarten on the achievement of students with low/moderate income status. Journal of
Research in Childhood Education, 20: 27-35.
Stofflet, F.P. (1998). Anchorage school district full-day kindergarten study: A follow-up of the kindergarten classes of 1987-88, 1988-89, and 1989-90.
Anchorage, AK: Anchorage School District, Kindergarten Experience Comparison (ERIC No. ED426790).
Uguroglu, M., Nieminen, G. (1986). Wilmette district #39 kindergarten study: Final report. Glen Ellyn, IL: The Institute for Educational Research (ERIC No.
ED294 681).
Walston, J., West, J. (2004). Full-day and half-day kindergarten in the United States: Findings from the early childhood longitudinal study, kindergarten class of
1998-99. Washington DC: U.S. Department of Education, National Center for Education Statistics. NCES 2004-078.
Winter, M., and Klein, A.E. (1970). Extending the kindergarten day: Does it make a difference in the achievement of educationally advantaged and
disadvantaged pupils? Washington, DC: Bureau of Elementary and Secondary Education (ERIC No. ED087534).
Wolgemuth, J.R., Cobb, R.B., Winokur, M.A., Leech, N., & Ellerby, D. (2006). Comparing longitudinal academic achievement of full-day and half-day
kindergarten students. Journal of Educational Research, 99(5): 260-269.

15
 Technical Appendices
Appendix A: Effect Size Procedures
A1: Study Selection and Coding Criteria
A2: Procedures for Calculating Effect Sizes
Appendix B: Methods and Parameters to Estimate the Benefits and Costs of Educational Options
B1: Valuation of Gains in Test Scores From Class Size Reductions and Full-Day Kindergarten
B2: Sensitivity/Risk Analysis
B3: The Per-Student Cost of Class Size Reductions
B4: The Per-Student Cost of Full-Day vs. Half-Day Kindergarten
Appendix C: Analysis of K–12 Outcomes
C1: Class Size Reduction Analysis
C2: Full-Day vs. Half-Day Kindergarten
Appendix A: Effect Size Procedures
This technical appendix describes the study coding criteria
and the procedures for calculating effect sizes that we use
in the Institute’s analysis of K–12 educational programs
and services. In recent years, researchers have developed
a set of statistical tools to facilitate systematic reviews of
evaluation evidence. The set of procedures is called
“meta-analysis” and we employ this methodology in our
study.28
A1. Study Selection and Coding Criteria
A meta-analysis is only as good as the selection and coding
criteria used to conduct the study. The following are key
coding criteria for our meta-analysis of evaluations of K–12
educational programs and services.
1) Study Search and Identification Procedures. We
search for all K–12 evaluation studies written in English.
We use three primary sources: a) study lists in other
reviews of the K–12 research literature; b) citations in
individual evaluation studies; and c) research
databases/search engines such as Google, Proquest,
Ebsco, ERIC, and SAGE.
2) Peer-Reviewed and Other Studies. Many K–12
evaluation studies are published in peer-reviewed
academic journals, while others are from government
or other reports. It is important to include non-peer
reviewed studies, because it has been suggested that
peer-reviewed publications may be biased toward
positive program effects. Therefore, our meta-
analysis includes studies regardless of their source.
3) Review of a Study’s Research Methodology. We
examine each potential study to ascertain whether the
study’s research design and data allow it to identify
causal effects of a program or policy on an
29
educational outcome. We include true experimental
studies and other non-experimental or observational
studies that have plausibly addressed the
endogeneity problem inherent in K–12 educational
studies. Econometric approaches to identify causal
effects include instrumental variables regression,
regression discontinuity designs, and fixed effects
panel models. Some multivariate correlational
designs employing hierarchical linear models,
ordinary least squares regression, and matching
16
designs are included if they have used a sufficient set
of right-hand side controls. We do not include studies
with a single-group, pre-post research design. We
believe that it is only through rigorous comparison
group studies that average treatment effects can be
30
reliably estimated.
4) Enough Information to Calculate an Effect Size.
Following the statistical procedures in Lipsey and Wilson
(2001), a study must provide the necessary statistical
information to calculate an effect size. If such
information is not provided, we attempt to contact the
author of the study. If this effort still does not produce
results, then we drop the study from our review.
5) Mean Difference Effect Sizes. For this study we are
coding mean difference effect sizes following the
procedures in Lipsey and Wilson (2001).
6) Unit of Analysis. Our unit of analysis is an
independent test of treatment at a particular site or
grade level. Some studies report outcome evaluation
information for multiple sites or grade levels; we
include each site or grade level as an independent
observation if a unique comparison group is also used
at each site.
7) Multivariate Results Preferred. Some studies
present two types of analyses: raw outcomes that are
not adjusted for covariates, such as family income
and ethnicity; and those that are adjusted with
multivariate statistical methods. In these situations,
we code the multivariate outcomes.
8) Some Special Coding Rules for Effect Sizes. Most
studies that meet the criteria for inclusion in our
review have sufficient information to code exact mean
difference effect sizes. Some studies report some,
but not all, of the information required. The rules we
follow for these situations are as follows:
a) Two-Tail P-Values. Sometimes, studies only
report p-values for significance testing of program
outcomes. If the study reports a one-tail p-value,
we will convert it to a two-tail test.
b) Declaration of Significance by Category. Some
studies report results of statistical significance
tests in terms of categories of p-values, such as
p<=.01, p<=.05, or “not significant at the p=.05
level.” We calculate effect sizes in these cases by
using the highest p-value in the category; e.g., if a
study reports significance at “p<=.05,” we
 calculate the effect size at p=.05. This is the most
conservative strategy. If the study simply states a
result was “not significant,” we compute the effect
size assuming a p-value of .50 (i.e. p=.50).
A2. Procedures for Calculating Effect Sizes
Effect sizes measure the degree to which a program has
been shown to change an outcome for program participants
relative to a comparison group. There are several methods
used by meta-analysts to calculate effect sizes, as
described in Lipsey and Wilson (2001). In this analysis, we
use statistical procedures to calculate standardized mean
difference effect sizes of programs. We do not use the
odds-ratio effect size because many of the outcomes
measured in this study, such as test scores, are
continuously measured.
A mean difference effect size involves continuous data
where the differences are in the means of an outcome.31
Mt − Mc
(A1) ESm =
SDt2 + SDc2
2
In this formula, ESm is the estimated effect size for the
difference between means obtained from the information in
a research study; Mt is the mean value of an outcome for
the treatment or experimental group; Mc is the mean value
of an outcome for the control group; SDt is the standard
deviation of the mean for the treatment group; and SDc is
the standard deviation of the mean for the control group.
Often, Mt - Mc is obtained from coefficients in regression
equations.
Some research studies report the mean values needed to
compute ESm in (A1), but they fail to report the standard
deviations. In these cases, if the authors report statistical
tests or confidence intervals, then this information allows
the pooled standard deviation to be estimated. These
procedures are described in Lipsey and Wilson (2001).
Some of the outcomes we record are measured as
dichotomies; for example, high school graduation. For these
yes/no outcomes, Lipsey and Wilson (2001) show that the
mean difference effect size calculation can be approximated
using the arcsine transformation of the difference between
32
proportions.
(A2) ES m = 2 × arcsin Pt − 2 × arcsin Pc
In this formula, ESm is the estimated effect size for the
difference between proportions from the research
information; Pt is the percentage of the population that had
an outcome for the experimental or treatment group; and Pc
is the percentage of the population that had an outcome for
the control or comparison group.
Adjusting Effect Sizes for Small Samples. Since some
studies have very small sample sizes, we follow the
recommendation of many meta-analysts and adjust for this.
Small sample sizes have been shown to upwardly bias
effect sizes, especially when samples are less than 20.
Following Hedges,33 Lipsey and Wilson34 report the “Hedges
correction factor,” which we use to adjust all mean
difference effect sizes (N is the total sample size of the
combined treatment and comparison groups):
⎡ 3 ⎤
(A3) ES′m = ⎢1 − ⎥ × [ES m ]
⎣ 4N − 9 ⎦
Adjusting Effect Sizes and Variances for Multi-Level
Data Structures. Most studies in the education field use
data that are hierarchical in nature. That is, students are
clustered in classrooms; classrooms are clustered in
schools; schools are clustered in districts; and districts are
clustered in states. These data structures require additional
adjustments.
There are two types of studies, each requiring a different
35
set of adjustments.
First, for child-level studies that ignore the variance due to
clustering, we make adjustments to the mean effect size
and its variance,
2(n − 1)ρ
(A4) EST = ES m ∗ 1 −
N −2
(A5)
⎛ N − Nc ⎞
V {EST } = ⎜⎜ t ⎟(1 + (n − 1)ρ ) + K
⎟
⎝ Nt Nc ⎠
⎛ (N − 2)(1 − ρ )2 + n(N − 2n )ρ 2 + 2(N − 2n )ρ (1 − ρ ) ⎞
K ES m 2 ⎜ ⎟
⎜ 2(N − 2)[(N − 2 ) − 2(n − 1)ρ ] ⎟
⎝ ⎠
where ρ is the intraclass correlation, the ratio of the variance
between clusters to the total variance; N is the total number
of individuals in the treatment group, Nt , and the
comparison group, Nc; and n is the average number of
persons in a cluster, K.
In the educational field, clusters can be classes, schools, or
districts. For this study, we used 2006 Washington
Assessment of Student Learning (WASL) data to calculate
values of ρ for the school-level (ρ = 0.114) and the district-
level (ρ = 0.052). Class-level data are not available for the
WASL, so we use a value of ρ = 0.200 for class-level
studies.
Second, for studies that report means and standard
deviations at a cluster level, we make adjustments to the
mean effect size and its variance:
1 + (n − 1)ρ
(A6) EST = ESm ∗ ∗ ρ
nρ
17
(A7) to .05, a random effects model is performed to calculate the
⎧⎪⎛ K − K ⎞ ⎛ 1 + ( n − 1) ρ ⎞ [1 + (n − 1) ρ ] ∗ ES 2 ⎫⎪ weighted average effect size. This is accomplished by first
V {EST } = ⎨⎜⎜ t c ⎟∗⎜ ⎟⎟ + m ∗ρ 42
⎟ ⎜ ⎬ calculating the random effects variance component, v.
⎪⎩⎝ t c ⎠ ⎝
K K n ρ ⎠ 2 nρ ( K t + K c − 2 )⎪
⎭
(A15) v = Qi − (k − 1)
∑ i (∑ wsqi ∑ wi )
w −

Computing Weighted Average Effect Sizes, Confidence

Intervals, and Homogeneity Tests. Once effect sizes are This random variance factor is then added to the variance of
calculated for each program effect, the individual measures are each effect size and finally all inverse variance weights are
summed to produce a weighted average effect size for a recomputed, as are the other meta-analytic test statistics.
program area. We calculate the inverse variance weight for
each program effect and these weights are used to compute
the average. These calculations involve three steps. First, the Appendix B: Methods and Parameters to
standard error, SET of each mean effect size is computed with:36 Estimate the Benefits and Costs of
Educational Options
Nt + N c ( EST )2
(A8) SET = +
Nt N c 2( Nt + N c ) This technical appendix describes our current model to
compute estimates of the benefits and costs of various
educational outcomes. Our approach employs a standard
Next, the inverse variance weight w is computed for each
37 human capital framework to value the outputs (effect
mean effect size with: sizes) of education inputs (e.g., class size reductions and
full-day kindergarten). Then, using other research that
1
(A9) w = has been conducted on the degree to which labor market
SET2 and other benefits accrue to those with improved
academic outcomes, we compute life-cycle benefits.
The weighted mean effect size for a group with i studies is Measuring the earnings implications of these human
computed with:38 capital variables in this manner is a commonly used
approach in economics.43 In recent years, economists
have also been estimating certain non-earnings outcomes
(A10) ∑ (wi EST )
ES = i
from indicators of improved education outcomes.44
∑ wi

Confidence intervals around this mean are then computed

B1. Valuation of Gains in Test Scores From Class
by first calculating the standard error of the mean with:39 Size Reductions and Full-Day Kindergarten
Exhibits B.1, B.2, and B.3 provide an illustration of how our
(A11) SE = 1
model computes benefits and costs of estimated gains in
ES
∑ wi standardized test scores. In this description, we use the
example of an estimated gain in test scores stemming from
Next, the lower, ESL, and upper limits, ESU, of the confidence a reduction in class size. We use the same procedures for
40
interval are computed with: our analysis of full-day kindergarten.

Column (3) in Exhibit B.3 indicates our estimated effect size,

(A12) ES L = ES − z(1−α ) ( SE ES )
(A13) ESU = ES + z(1−α ) ( SE ES )
In equations (A8) and (A9), z(1-α) is the critical value for the
z-distribution (1.96 for α = .05).
The test for homogeneity, which provides a measure of
the dispersion of the effect sizes around their mean, is
given by:41
(∑ wi ESi ) 2
(A14) Qi = (∑ wi ESi2 ) −
∑ wi
The Q-test is distributed as a chi-square with k-1 degrees of
freedom (where k is the number of effect sizes).
Computing Random Effects Weighted Average Effect
Sizes and Confidence Intervals. When the p-value on the
Q-test indicates significance at values of p less than or equal
in standard deviation units of a standardized test score, in
the grade in which the class size reduction takes place. In
the example shown in Exhibit B.3, we have estimated the
results for a class size reduction of one unit during first
grade. The input parameters are shown in Exhibit B.1. We
then use another parameter to model any expected annual
rate of decay (or growth) in this effect size by the end of high
school. This adjustment to align effect sizes in an early
grade with effect sizes in later grades is made because the
long-run effect of improved test scores on earnings has
generally been estimated by economists for high school test
scores. Equation (B1) describes this process, where ES18 is
the estimated effect size at age 18. It is calculated with the
effect size age of the student during the program year,
ESprogyear (first grade in our example), and an annual rate of
decay in the effect size, ESdecay.
(B1) ES18 = ES progyear × (1 + ESdecay)18− progyear

18
In our analysis, all human capital earnings estimates derive
from a common dataset. The estimates are taken from the
U.S. Census Bureau’s March Supplement to the Current
Population Survey, which provides cross-sectional data for
earnings by age and by educational status. To these data,
we apply different measures of the net advantage gained
through increases in a human capital outcome such as test
scores. The level of earnings shown on column (4) of Exhibit
B.3 is taken from cross-sectional data from the 2005 March
Supplement to the Current Population Survey (CPS), with
45
data on earnings during 2004. The earnings are those for
people with education levels between 9th grade through some
college. The number of non-earners is included in the
estimates so that the average earning level reflects earnings
of all people at each age (earners and non-earners).
In column (5), we adjust these CPS earnings data for general
inflation to bring the CPS data, denominated in 2004 dollars,
up to the base year for our analysis (2006), fringe benefits,
and the economy-wide real growth rates in earnings.
Equation B2 describes this adjustment process.
× (1 + Fringe) × (1 + Earnesc) y −18 health care costs; reduced crime; reduced foster care; and
IPDbase
(B2) Earn y = CPSEarn y ×
IPDcps
CPS earnings in each year (from age 18 to age 65),
CPSEarny, are first converted to 2006 dollars with an
inflation index. The inflation index is taken from the
Washington State Economic and Revenue Forecast
Council, the official forecasting agency for Washington
State government. The index is the chain-weight implicit
price deflator for personal consumption expenditures. In 46
equation B2, this adjustment is IPDbase / IPDcps.
We then adjust for an estimate of the average fringe benefit
rate for earnings, Fringe. This estimate is from the
Employment Cost Index as computed by the U.S. Bureau of
Labor Statistics.47
We also adjust for long-run expected growth rates in real
earnings, Earnesc. The estimate for the medium case is
taken from the Congressional Budget Office (CBO)
48
analysis of long-run Social Security. We model a higher
rate of growth and a lower rate of growth in our sensitivity
analyses (ranges shown on Exhibit B.2).
In column (6) of Exhibit B.3, we indicate the gain in
earnings with a one standard deviation increase in test
scores each year. In equation B3, this is given by
OneSDEarny.
(B3) OneSDEarny = Earn y × TSROR × (1 + TSROResc) y −18
In this equation we multiply the earnings estimates from B2
by an estimate of the rate of return on earnings from a one
standard deviation increase in test scores, TSROR. Our
estimate of this factor follows the summary made by
Hanushek (2004) of recent economic analyses (our
49
estimates are shown on Exhibit B.2). Hanushek (2004)
also describes economic research indicating that the
expected rate of return from test scores on earnings may
grow over time as the market in general, and employers in
particular, place increasingly higher values on skills and
schooling.50 We provide for this in equation B3 by including
an estimate for the annual rate of escalation in the rate of
return to test scores, TSROResc. We also test various rates
for these factors in our sensitivity analyses (ranges shown
on Exhibit B.2).
In column (7), we show estimates for the annual earnings
gained from the example increase in test scores. These
amounts are estimated with equation B4, where adjusted
earnings, AdjEarny, in each year is derived from the
expected gain in earnings from a one standard deviation
gain multiplied by the end-of-high school estimated effect
size (also in standard deviation test score units).
(B4) AdjEarn y = OneSDEarn y × ES18
Recent research literature has also focused attention on
several types of non-market or social benefits associated,
perhaps causally, with human capital education outcomes.
A listing of possible non-market benefits to education
51
appears in the work of Wolfe and Haveman, and Riddell.
These factors include “knowledge spillovers” that stimulate
general economic growth; improved health care and lower
increased civic participation. In our current cost-benefit
model, we provide a simple multiplicative parameter that
can be applied to the estimated earnings effects so that
the non-market benefits can be roughly modeled. Since
some research indicates that these non-market benefits of
human capital outcomes can be considerable, future
refinements to our cost-benefit model will attempt to
analyze these possible non-wage benefits explicitly. In the
meantime, we run our model with and without the social
benefits included, and we test various assumed levels of
social rates of return.
Our model calculates these non-market benefits using a
social rate of return parameter as applied to the earnings
estimates already discussed. These are shown in columns
(8) and (9) of Exhibit B.3 and are described with equations
B5 and B6. As before, earnings are multiplied by an
assumed social rate of return, SocialROR, and then the
resulting series is multiplied by the effect size at age 18.
High and low ranges for this social return factor are
modeled in a simulation framework (ranges shown on
Exhibit B.2).
(B5) OneSDSocial y = Earn y × SocialROR
(B6) AdjSocial y = OneSDSocial y × ES18
Present values of the estimated market and total benefits are
then computed with the information in columns (7) and (9) of
Exhibit B.3. For example, for the total benefits case, equation
B7 first discounts the sum of the labor market earnings and
the social earnings to age 18, with a real discount rate, Dis.
Equation B8 then discounts this sum further to the year in
which the investment in the K–12 resources are made, page
(e.g. the age of the student in the grade level when the
resources were spent to lower class sizes).
65 AdjEarn + AdjSocial
∑
y y
(B7) PVBen18 =
y −18
y =18 (1 + Dis )
19
 PVBen18
(B8) PVBen page =
18− progyear
(1 + Dis )
We estimate a range of real discount rates for this study.
The high end of the range is a 7 percent real discount rate.
This discount rate reflects the rate that has been
recommended by the federal Office of Management and
52
Budget. The low end of the range is a 3 percent real
discount rate used by the Congressional Budget Office in
a variety of analyses including its projections of the long-
term financial position of Social Security.53 Our study uses
a medium real discount of 5 percent, the midpoint between
54
the high and low rates.
Finally, in columns (10) and (11) of Exhibit B.3, we
arrange the annual cash or resource flows to enable the
calculation of internal rates of return on investment. The
average cost per student to lower class size, discussed
below, is placed in the year in which the resources are
spent, and the benefits from equation B8 are arranged
accordingly. An internal rate of return for this stream of
cash and resource flows is then computed with Microsoft
Excel’s IRR function.
B2. Sensitivity/Risk Analysis
The model as described in this appendix produces a
unique result given the set of inputs listed. As we
describe, however, there is uncertainty around many of
the inputs. For most inputs to the model, we determine
the range of uncertainty with the standard errors or
standard deviations from relevant statistics of the
underlying data for each parameter. For a few other
parameters, we hypothesize low and high ranges to place
bounds on our estimates of uncertainty.
After we specify ranges of uncertainty on each of the
inputs, we then use a simulation approach to determine
the degree to which the final result is sensitive to these
known or hypothesized levels of uncertainty. To conduct
the simulation, we use Palisade Corporation’s @RISK®
simulation software. Using a Monte Carlo approach for
the simulation, the software randomly draws from the
user-designated input variables after a particular type of
probability distribution and its parameters have been
specified for the input. We run a Monte Carlo simulation
for 5,000 cases. Exhibit B.2 shows the range of variability
for the key input variables we use in the simulations.
B3. The Per-Student Cost of Class Size
Reductions
The calculation of costs and benefits requires an estimate
of the taxpayer cost of reducing class sizes. We provide
our estimates in Exhibit B.4. We estimate operating and
capital costs associated with unit changes in the number
of students per classroom. The cost estimate is driven by
the following six parameters, shown at the bottom of
Exhibit B.4:
1) Average annual teacher salary in an average classroom
(non-wage benefits included, 2006 dollars);
2) Total number of public K–12 students in Washington (or
any given grade and geographic cohort);
3) Average square feet of K–12 classroom per student;
20
4) Construction cost for K–12 classrooms (dollars per
square foot, 2006 dollars);
5) Length of bonds for new construction; and the
6) Interest rate on bonds.
The operating cost estimates in columns (2) and (3) of
Exhibit B.4 are simply the level and change in per-student
teacher expenses as class size changes from one level to
the next. The capital cost calculations in columns (4)
through (8) begin by estimating the number of classrooms
needed, and the change in the number of classrooms
needed, as class size changes by one unit for a given
population (in our example, we estimate it for the entire
number of students in the public K–12 system). We then
multiply this by the change in the number of classrooms
from one average class size to the next, and then by the
number of square feet per average classroom and the cost
per square foot of new construction. This product is then
financed over an assumed bond term and interest rate.
The result is then divided by the student population to
estimate a per-student capital cost.
The per-student operating and capital costs are combined
in Exhibit B.4 in column (9) to provide an estimate of the
total per-student cost of reducing class size from one level
to one level smaller. For example, the per-student cost of
reducing class size from 20 to 19 is $236 per student. The
cost of reducing class size from 23 to 15 would be $2,054
($354+$317+$286+$259+$236+$217+$200+$185).
B4. The Per-Student Cost of Full-Day vs. Half-
Day Kindergarten
We provide an estimate of the average per-student cost of
moving from half-day to full-day kindergarten in Exhibit
B.5. We calculate operating and capital costs. The cost
estimate is driven by the following seven parameters,
shown at the bottom of Exhibit B.5:
1) Average annual teacher salary in an average classroom
(non-wage benefits included, 2006 dollars);
2) Total number of public kindergarten students in
Washington (or any geographic sub-unit);
3) Average kindergarten students per classroom;
4) Average square feet per average K–12 classroom;
5) Construction cost for K–12 classrooms (dollars per
square foot, 2006 dollars);
6) Length of bonds for new construction; and the
7) Interest rate on bonds.
The difference in operating costs is estimated as simply
the difference in average teacher salary for an FTE
teacher, given an average kindergarten class size. This
estimate does not include any estimated effects on pupil
transportation costs of moving from half-day to full-day
kindergarten. The capital cost calculations estimate the
number of additional classrooms needed, times the
number of square feet per student, and the cost per
square foot of new construction. This product is then
financed over an assumed bond term and interest rate.
The result is then divided by the student population to
estimate a per-student capital cost.
 Exhibit B.1
Example Calculation: Input Parameters and
Return on Investment Results for Class Size Reduction
Parameters for the Calculation of Return on Investment
1 Grade for which the gain in test scores is initially estimated
0.019 Initial gain (standard deviation units on test scores) for this grade level
-0.080 Annual rate of decay or growth in effect size (from initial effect to end of high school)
0.013 Average annual real rate of growth in earnings
0.423 Fringe benefit percentage for earnings
1.057 Inflation rate, 2004 to 2006, Implicit Price Deflator
0.118 Percent change in annual earnings with a one standard deviation gain in test scores
0.005 Annual real rate of growth for this return on test score change percentage
0.070 Social rate of return (as a function of labor market earnings)
0.050 Real discount rate for the analysis
21 Average class size before class size reduction
20 Average class size after class size reduction
Return on Investment Labor Market Only Total Return
Internal rate of return on investment 7.0% 8.7%
Present value of the benefits, per student $378 $581
Per student cost for the class size reduction (operating and capital) $217 $217
Net present value per student (benefits minus costs) $162 $364
Benefit-to-cost ratio $1.75 $2.68

Exhibit B.2
Sensitivity/Risk Analysis for the Economic Model
Parameters for the Calculation of Return on Investment
0.019 Initial gain (standard deviation units on test scores) for this grade level from K–12 program or policy
0.003 Standard error
-0.080 Annual rate of decay or growth in effect size (from initial effect to end of high school)
0.000 Minimum decay rate
-0.160 Maximum growth rate
0.013 Average annual real rate of growth in labor market earnings
0.023 High rate
0.003 Low rate
0.118 Percent change in annual earnings with a one standard deviation gain in test scores
0.018 Standard error
0.005 Annual real rate of growth for this economic return on test scores
0.010 High rate
0.000 Low rate
0.070 Social rate of return (as a function of labor market earnings)
0.100 High rate
0.000 Low rate
0.050 Real discount rate for the analysis
0.070 High rate
0.030 Low rate
21 Average class size before class size reduction
20 Average class size after class size reduction
1 Grade for which the gain in test scores is initially estimated
0.423 Fringe benefit percentage for earnings
1.057 Inflation rate, 2004 to 2006, Implicit Price Deflator

21
 Exhibit B.3
Worksheet to Estimate Return on Investment for Programs and Policies
That Increase Standardized Test Scores
Test Score Change Labor Market Earnings Change Other Gains Summary

Effect Size: Average Average Gain in

Standard earnings, earnings with earnings Earnings Gain in other Other
deviation gain in workers and non- fringe benefits from a one gain, end- benefits from gains, end- Total
standardized test workers, Current and real standard of-high a one of-high Annual annual cash
scores for the Population growth in deviation school standard school labor and
Age of K-12 class size Survey, 2004 earnings, 2006 gain in test effect of deviation gain effect of market resource
person grade reduction dollars dollars scores test score in test scores test score cash flows flows
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
1 -4
2 -3
3 -2
4 -1
5 0
6 1 0.01943 $0 $0 -$217 -$217
7 2 0.01788 $0 $0 $0 $0
8 3 0.01645 $0 $0 $0 $0
9 4 0.01513 $0 $0 $0 $0
10 5 0.01392 $0 $0 $0 $0
11 6 0.01281 $0 $0 $0 $0
12 7 0.01178 $0 $0 $0 $0
13 8 0.01084 $0 $0 $0 $0
14 9 0.00997 $0 $0 $0 $0
15 10 0.00918 $0 $0 $0 $0
16 11 0.00844 $0 $0 $0 $0
17 12 0.00777 $0 $0 $0 $0
18 - $3,174 $4,775 $565 $4 $334 $3 $4 $7
19 - $5,741 $8,749 $1,040 $8 $612 $5 $8 $13
20 - $7,972 $12,308 $1,470 $11 $862 $7 $11 $18
21 - $10,316 $16,132 $1,937 $15 $1,129 $9 $15 $24
22 - $11,527 $18,261 $2,203 $17 $1,278 $10 $17 $27
23 - $14,325 $22,988 $2,788 $22 $1,609 $12 $22 $34
24 - $15,325 $24,913 $3,036 $24 $1,744 $14 $24 $37
25 - $18,032 $29,694 $3,637 $28 $2,079 $16 $28 $44
26 - $18,144 $30,267 $3,726 $29 $2,119 $16 $29 $45
27 - $19,968 $33,743 $4,174 $32 $2,362 $18 $32 $51
28 - $20,505 $35,101 $4,364 $34 $2,457 $19 $34 $53
29 - $22,468 $38,961 $4,868 $38 $2,727 $21 $38 $59
30 - $22,530 $39,577 $4,970 $39 $2,770 $22 $39 $60
31 - $24,514 $43,622 $5,505 $43 $3,054 $24 $43 $66
32 - $23,978 $43,222 $5,482 $43 $3,026 $23 $43 $66
33 - $22,431 $40,960 $5,221 $41 $2,867 $22 $41 $63
34 - $23,354 $43,198 $5,534 $43 $3,024 $23 $43 $66
35 - $25,804 $48,351 $6,225 $48 $3,385 $26 $48 $75
36 - $27,221 $51,670 $6,685 $52 $3,617 $28 $52 $80
37 - $26,220 $50,417 $6,556 $51 $3,529 $27 $51 $78
38 - $26,894 $52,386 $6,846 $53 $3,667 $28 $53 $82
39 - $27,028 $53,329 $7,004 $54 $3,733 $29 $54 $83
40 - $27,636 $55,240 $7,291 $57 $3,867 $30 $57 $87
41 - $27,153 $54,979 $7,293 $57 $3,849 $30 $57 $87
42 - $27,214 $55,819 $7,441 $58 $3,907 $30 $58 $88
43 - $28,534 $59,287 $7,943 $62 $4,150 $32 $62 $94
44 - $28,222 $59,402 $7,998 $62 $4,158 $32 $62 $94
45 - $28,414 $60,582 $8,198 $64 $4,241 $33 $64 $97
46 - $27,974 $60,420 $8,217 $64 $4,229 $33 $64 $97
47 - $27,794 $60,812 $8,312 $65 $4,257 $33 $65 $98
48 - $28,189 $62,478 $8,582 $67 $4,373 $34 $67 $101
49 - $28,038 $62,951 $8,690 $67 $4,407 $34 $67 $102
50 - $27,896 $63,445 $8,802 $68 $4,441 $34 $68 $103
51 - $27,865 $64,200 $8,952 $70 $4,494 $35 $70 $104
52 - $28,098 $65,578 $9,190 $71 $4,590 $36 $71 $107
53 - $25,713 $60,791 $8,561 $66 $4,255 $33 $66 $100
54 - $26,649 $63,824 $9,033 $70 $4,468 $35 $70 $105
55 - $26,356 $63,943 $9,096 $71 $4,476 $35 $71 $105
56 - $23,163 $56,926 $8,138 $63 $3,985 $31 $63 $94
57 - $25,921 $64,533 $9,271 $72 $4,517 $35 $72 $107
58 - $21,941 $55,335 $7,990 $62 $3,873 $30 $62 $92
59 - $22,215 $56,753 $8,235 $64 $3,973 $31 $64 $95
60 - $23,097 $59,775 $8,717 $68 $4,184 $32 $68 $100
61 - $19,166 $50,247 $7,364 $57 $3,517 $27 $57 $85
62 - $17,390 $46,181 $6,802 $53 $3,233 $25 $53 $78
63 - $12,120 $32,605 $4,827 $37 $2,282 $18 $37 $55
64 - $11,068 $30,162 $4,487 $35 $2,111 $16 $35 $51
65 - $8,034 $22,179 $3,316 $26 $1,552 $12 $26 $38

22
 Exhibit B.4
Estimated Annual Per-Student Cost of Lowering K–12 Class Size
Operating Costs Capital Costs Total Cost

Change in the Number of Change in the Change in the Annual capital Total annual
salary cost per classrooms number of square footage amortization cost per
Salary cost per student for a one needed for a classrooms for a for a one unit costs for a one Annual student for a
Class Size student for a unit drop in given one unit drop in drop in unit drop in capital one unit drop
(students per given average class classroom average class average class average class payment in average
classroom) classroom size size size size size size per student class size
(1) (2) (3) (4) (5) (6) (7) (8) (9)
10 $6,690 $608 102,731 9,339 8,405,280 $112,789,473 $110 $718
11 $6,082 $507 93,392 7,783 7,704,840 $103,390,351 $101 $607
12 $5,575 $429 85,609 6,585 7,112,160 $95,437,247 $93 $522
13 $5,146 $368 79,024 5,645 6,604,149 $88,620,300 $86 $454
14 $4,779 $319 73,379 4,892 6,163,872 $82,712,280 $81 $399
15 $4,460 $279 68,487 4,280 5,778,630 $77,542,763 $75 $354
16 $4,181 $246 64,207 3,777 5,438,711 $72,981,424 $71 $317
17 $3,935 $219 60,430 3,357 5,136,560 $68,926,900 $67 $286
18 $3,717 $196 57,073 3,004 4,866,215 $65,299,169 $64 $259
19 $3,521 $176 54,069 2,703 4,622,904 $62,034,210 $60 $236
20 $3,345 $159 51,366 2,446 4,402,766 $59,080,200 $58 $217
21 $3,186 $145 48,920 2,224 4,202,640 $56,394,737 $55 $200
22 $3,041 $132 46,696 2,030 4,019,917 $53,942,792 $53 $185
23 $2,909 $121 44,666 1,861 3,852,420 $51,695,175 $50 $172
24 $2,788 $112 42,805 1,712 3,698,323 $49,627,368 $48 $160
25 $2,676 $103 41,092 1,580 3,556,080 $47,718,623 $46 $149
26 $2,573 $95 39,512 1,463 3,424,373 $45,951,267 $45 $140
27 $2,478 $88 38,049 1,359 3,302,074 $44,310,150 $43 $132
28 $2,389 $82 36,690 1,265 3,188,210 $42,782,214 $42 $124
29 $2,307 $77 35,425 1,181 3,081,936 $41,356,140 $40 $117
30 $2,230 $72 34,244 1,105 2,982,519 $40,022,071 $39 $111
31 $2,158 $67 33,139 1,036 2,889,315 $38,771,381 $38 $105
32 $2,091 $63 32,104 973 2,801,760 $37,596,491 $37 $100
33 $2,027 $60 31,131 916 2,719,355 $36,490,712 $36 $95
34 $1,968 $56 30,215 863 2,641,659 $35,448,120 $35 $91
35 $1,911 $53 29,352 815 2,568,280 $34,463,450 $34 $87
36 $1,858 $50 28,536 771 2,498,867 $33,532,006 $33 $83
37 $1,808 $48 27,765 731 2,433,107 $32,649,584 $32 $79
38 $1,761 $45 27,035 693 2,370,720 $31,812,416 $31 $76
39 $1,715 $43 26,341 659 2,311,452 $31,017,105 $30 $73
40 $1,673 - 25,683 -

Assumed Parameters in Cost Calculation

$66,900 Average annual teacher salary in an average classroom (non-wage benefits included, 2006 dollars)
1,027,312 Total number of public K-12 students in Washington
90 Average square feet of classroom space per student
$180 Construction cost for K-12 classrooms (dollars per square foot, 2006 dollars)
25 Length of bonds for new construction
5.50% Interest rate on bonds

Exhibit B.5
Estimated Annual Per-Student Cost of Full-Day vs. Half-Day Kindergarten
Half-Day K Full-Day K Difference
72,824 72,824 Students in cohort
0.5 1.0 Full time equivalent teacher given to each student
20 20 Average kindergarten class size
1,821 3,641 Number of teachers needed, FTE
$2,007.00 $4,014.00 Teacher cost per student (includes marginal non-teacher salary operating expenses)
$2,007.00 Difference in operating cost per student
1,821 3,641 Number of classrooms needed
3,277,080 6,554,160 Total square footage of classrooms
3,277,080 Change in square footage
$589,874,400 Construction cost for change in square footage
$43,974,755 Annual payment to capital
$603.85 Capital payment per student
$2,610.85 Total cost per student to expand from half-day to full-day kindergarten

Assumed Parameters in Cost Calculation

$66,900 Average annual teacher salary in an average classroom (non-wage benefits included, 2006 dollars)
20% Marginal non-teacher salary operating expenses (as percent of teacher salaries)
72,824 Total number of public kindergarten students in Washington
20 Average kindergarten class size
90 Average square feet of classroom space per student
$180 Construction cost for K-12 classrooms (dollars per square foot, 2006 dollars)
25 Length of bonds for new construction
5.50% Interest rate on bonds

23
Appendix C: Analysis of K–12 Outcomes while the constant term in the regression is the estimated
effect of the omitted kindergarten to grade 2 group. The
effect for the K to grade 2 group is .0194 standard deviation
C1. Class Size Reduction Analysis units per one-unit reduction in class size and the result is
statistically significant (p=.000). The coefficient test
Multivariate Results. As described in the main section of (C1)+(C2)=0 indicates the effect of a one-unit reduction in
this report, we found 38 mostly recent high-quality studies class size in grades 3 through 6 is .007 standard deviation
examining the effect of class size reductions on academic units and this result is significantly different from zero
outcomes. These studies contained 69 separate effect (p=.000). The result for grades 7 through 8, coefficient test
sizes, where an effect size is the change in standard (C1)+(C3)=0, indicates a -.001 reduction in standard
deviation units for a one-unit change in class size. There deviation units but the result is not significant (p=.555).
are more effect sizes than studies because some of the Finally, the result for grades 9 through 12 is .004 standard
studies estimated outcomes for multiple locations or deviation units per one-unit reduction in class size; this
multiple grades, for separate populations. We performed result is also not significant (p=.179).
multivariate analysis on these effect sizes in order to
produce estimates of mean effects, by grade level, along Because the sample sizes of the studies used in this
with statistical significance tests. We estimated many analysis vary so widely, the inverse variance weights also
models with different sets of covariates and with different vary widely (minimum=46; maximum=382,789;
weighting series for use with weighted least squares average=13,141; median=427). We estimated different
regression. We describe some of these additional tests models where a maximum cutoff level for the weights was
below. Our preferred model is the parsimonious one imposed so that any study with an inverse variance greater
shown in Exhibit C.1. than the cutoff level was assigned the cutoff level weight.
Because of the wide dispersion in weights, when no cutoff
Exhibit C.1 level is imposed, the regression is dominated by just a few
Preferred Regression Model studies. At the other extreme, when no inverse variance
Dependent Variable: ESTOT weights are used, then each study carries a weight of one,
Method: Least Squares meaning that the smallest study is given equal weight with
Included observations: 69 studies that have substantial sample sizes. Both of these
Weighting series: INVTOT500
extremes are less than optimal, so we selected a maximum
White Heteroskedasticity-Consistent Standard Errors &
Covariance
cutoff value around the median inverse variance weight
(500 in our preferred model), and then tested some larger
cutoff levels for sensitivity. Exhibit C.2 shows the results of
Variable Coefficient Std. Error t-Statistic Prob.
Constant (C1) 0.019434 0.003493 5.56 0.000
models with different weighting series. The coefficients
Grades 3 to 6 (C2) -0.012712 0.003774 -3.36 0.001 and their significance are quite stable except in the extreme
Grades 7 to 8 (C3) -0.020780 0.004172 -4.98 0.000 case of no restrictions on the inverse variance weights
Grades 9 to 12 (C4) -0.015831 0.004407 -3.59 0.000 where the effect for grades 3 to 6 drops to zero; this case
R-squared 0.418 Mean dep. var 0.0072 also has an implausibly large adjusted R-square, .988,
Adjusted R-squared 0.391 S.D. dependent var 0.0130 indicating that the regression was adjusting for only one or
S.E. of regression 0.0108 Akaike criterion -6.284 two large studies.
Sum squared resid 0.0067 Schwarz criterion -6.1554
Log likelihood 220.8 F-statistic 13.410
Durbin-Watson stat 1.76 Prob (F-statistic) 0.000
Exhibit C.2
Tests of the Regression Model in Exhibit C.1
Coefficient Tests Coefficient Std. Error t-Statistic Prob.
With Different Inverse Variance Weighting Series
(C1)+(C2)=0 0.006723 0.001428 4.70 0.000 for the Weighted Least Squares Regression
(C1)+(C3)=0 -0.001346 0.002282 -0.59 0.555 (regression coefficients with standard errors in parentheses)
(C1)+(C4)=0 0.003603 0.002687 1.34 0.179
Variable No Max = Max = Max = No
Weights 500 1,000 10,000 Restriction
The model is a weighted least squares regression where 0.019 0.019 0.019 0.018 0.018
Constant
the weights are the inverse variances, adjusted for (0.002) (0.003) (0.003) (0.003) (0.003)
clustering, calculated for each effect size as described in Grades 3 to 6
-0.011 -0.013 -0.012 -0.010 -0.003
Appendix A. The dependent variable, ESTOT, is the (0.003) (0.004) (0.003) (0.003) (0.003)
average test score measured in each study. If a study -0.019 -0.021 -0.019 -0.020 -0.023
Grades 7 to 8
(0.005) (0.004) (0.003) (0.003) (0.003)
supplied a math score and a reading score, then we
-0.014 -0.016 -0.015 -0.015 -0.010
averaged the two effect sizes to produce an average test Grades 9 to 12
(0.003) (0.004) (0.004) (0.005) (0.004)
score effect size for that study. Separately, we also Adj R-Square 0.240 0.391 0.478 0.544 0.988
conducted regressions just on those studies with math
scores and reading scores.
We also tested a number of other covariates in a variety of
We created dummy variables for grades 3 through 6, 7 model structures. Exhibit C.3 shows the results for our
through 8, and 9 through 12; the omitted category in the basic model with the addition of covariates describing
regression was kindergarten through grade 2. We tested attributes of the studies. None of the covariates is
different grade-level groupings, particularly in the early statistically significant, and the coefficients on the policy
grades; this combination was representative of the results. variables remain quite close to those in the parsimonious
Because this parsimonious model has no other covariates, model. We create a dummy variable, IDMETHOD, for
the coefficient tests shown in Exhibit C.1 are the marginal those studies that are either random assignment studies,
effects of a one-unit reduction in class size for each group, instrumental variables studies, or regression discontinuity

24
studies (74 percent of the 69 effect sizes) and these Long-Term Decay of Test Score Gains. For the most
studies were coded one; correlational studies (hierarchical part, each of the 69 effect sizes in our study analyzed the
linear models or ordinary least squares) were coded zero. results of a standardized test administered quite close to
This variable was not significant (p=.83). the time when class sizes were reduced. Therefore, the
results that we estimate with our preferred regression
Studies based on a population in the U.S. (49 percent of model should be regarded as near-term changes in test
the 69 effect sizes) were coded one; international studies scores for a one-unit change in class size. An open
were coded zero. This variable was not significant (p=.98). question concerns whether these effect sizes decay over
Studies based on student-level data (77 percent of the 69 time. That is, if a class size reduction induces a test score
effect sizes) were coded one; studies based on class, gain in first grade, does that effect size maintain throughout
school, or district data were coded zero. This variable was the K–12 years? This question is important, because the
not significant (p=.23). Studies based on the policy economic model described in Appendix B indicates that
variable “total spending” (10 percent of the 69 effect sizes) long-term labor market and other benefits accrue to gains
were coded one; studies based on the policy variable class in test scores in the upper grades.
size were coded zero. This variable was not significant
(p=.84). Finally, studies based on a dependent variable Unfortunately, only a few of the studies we reviewed for this
that was not a test score but, instead, used high school report contain long-term follow-up information on subsequent
graduation as the outcome (3 percent of the 69 effect effect sizes. And, the results of the studies that do have follow-
sizes) were coded one; studies based on a dependent that up data suggest inconsistent findings. The primary class-size
was a test score were coded zero. This variable was not study that examined longer-term results is the Tennessee
significant (p=.40). STAR experiment. Nye et al. (1999)55 and Nye et al. (2001)56
followed the K–3 STAR students into middle school and found
Exhibit C.3 virtually no decay in effects in subsequent test scores. On the
57
Preferred Regression Model With Covariates other hand, Krueger and Whitmore (2001) studied whether
Dependent Variable: ESTOT the STAR students took college placement tests in high school;
Method: Least Squares they found that STAR students had a statistically higher
Included observations: 69 chance of taking the test (43.7 percent compared with 40.0
Weighting series: INVTOT500 percent of the comparison group). This effect size, however, is
White Heteroskedasticity-Consistent Standard Errors & just .041 (see the arcsine transformation listed in equation
Covariance
(A2)) compared with the original test score effect size Krueger
found during grades K through 3 (about .22). Thus, in terms of
Variable Coefficient Std. Error t-Statistic Prob.
effect sizes that measure academic success, Krueger’s effect
Constant (C1) 0.016066 0.005674 2.831294 0.0063
Grades 3 to 6 (C2) -0.011045 0.004008 -2.755583 0.0077
size declined at about a 16 percent annual rate of decay
Grades 7 to 8 (C3) -0.021097 0.003576 -5.900371 0.0000 between the early grades and high school. Another long-term
Grades 9 to 12 (C4) -0.016613 0.004078 -4.073533 0.0001 effect of the STAR experiment was measured by Finn et al.
IDMETHOD -0.000757 0.003591 -0.210907 0.8337 (2005).58 They examined high school graduation rates and
USA 0.000101 0.003893 0.026036 0.9793 found that students who did not spend any time in the smaller
STUDENTLEVEL 0.004699 0.003869 1.214463 0.2293 STAR classes had a 76.3 percent high school graduation rate,
SPENDSTUDY 0.000975 0.004770 0.204311 0.8388
while the STAR students averaged a 79.6 percent graduation
NOTTESTSCORE 0.004025 0.004758 0.845978 0.4009
rate—an effect size of .053 (again, with the arcsine
R-squared 0.450093 Mean dep. var 0.0072 approximation). This effect size is similar to the long-term rate
Adjusted R-squared 0.376772 S.D. dependent var 0.0130
0.010275
found by Krueger and is considerably below the level of effect
S.E. of regression Akaike criterion -6.197
Sum squared resid 0.006334 Schwarz criterion -5.9057 sizes for test scores in the early grades.
Log likelihood 222.8015 F-statistic 5.356
Durbin-Watson stat 1.787983 Prob (F-statistic) 0.000 Because of these inconsistent results, we have modeled a
variety of long-term effect size decay parameters in our
In other analyses (not shown), we tested whether the initial economic models. As shown in Exhibit B.2 in our simulation
class size level, before the class size reduction, was models, we model a 16 percent annual rate of effect size
significant. It was never close to significant in any of the decay for the high decay case; a zero percent rate of decay for
models we tested, with p-values around .50. the low case; and an averaged 8 percent rate of decay for the
medium case. Clearly, more research needs to be performed
As mentioned, the dependent variable in the models on the long-term effects of class size reductions.
presented is the average effect size for each of the 69
separate trials. Some of the trials had two or more test Analysis of Effects of Class Size Changes on Low-Income
score results. For example, some trials had a math test Populations. Only one of the 69 class size studies reviewed
score result and a reading test score result administered to for this study specifically examined the interaction between
59
the same group of students who received the same class size and student low-income status. However, nine
reduction in class size. In our general models, we studies (reporting 18 separate effect sizes) provided
averaged these test score effect sizes to calculate an information on the low-income status of students in their study
average effect size for each of the 69 trials. Other trials populations (i.e., the proportion with free or reduced lunch).
just reported a general test score result. We ran models The studies used in this analysis, their effect sizes as we
(not shown) for just the trials that had math tests and, calculated with the methods in Appendix A, the percentage of
separately, for those with reading, writing, or language low-income students, and the grade at intervention are
tests. We found results to be consistent with the findings in described in Exhibit C.4.
our parsimonious preferred model.

25
 Exhibit C.4 Exhibit C.5
Data From Studies Describing the Proportion Linear Model
of Students in Low-Income Families Dependent Variable: ESTOT
Study Average Percent Method: Least Squares
Grade
Effect Size Low-income Included observations: 18
Akerhielm (1995) 0.0034 39.3 8 Weighting series: INVTOT500
Ecalle et al (2006) 0.0137 24.0 1
Feinstein & Symons (1999) 0.0038 25.0 10
Hoxby (2000) 0.0014 24.9 2 Variable Coefficient Std. Error t-Statistic Prob.
Hoxby (2000) -0.0007 24.9 3 Constant -0.00224 0.00702 -.32 0.7538
LowSES 0.004821 0.000125 3.87 0.0015
Hoxby (2000) 0.0001 24.9 2
Grade -0.000726 0.000653 -1.11 0.2834
Hoxby (2000) 0.0000 24.9 4
R-squared 0.7151 Mean dep. var 0.0137
Krueger (1999) 0.0275 47.0 0
Adjusted R-squared 0.6771 S.D. dependent var 0.0124
Krueger (1999) 0.0404 59.0 1 S.E. of regression 0.14131 F-statistic 18.82
Krueger (1999) 0.0266 66.0 2 Sum squared resid 0.75166 Prob (F-statistic) 0.0001
Krueger (1999) 0.0229 60.0 3
Levacic et al (2005) -0.0019 13.0 8
Levacic et al (2005) 0.0055 11.8 9.5 Exhibit C.6
Molnar et al (1999) 0.0196 57.7 1 Curvilinear Model
Molnar et al (1999) 0.0173 54.0 1 Dependent Variable: ESTOT
NICHD (2004) 0.0174 32.2 1 Method: Least Squares
Sander (1999) 0.0014 20.9 8 Included observations: 18
Sander (1999) 0.0022 23.0 3 Weighting series: INVTOT500

We performed multivariate analyses on these effect sizes

Variable Coefficient Std. Error t-Statistic Prob.
in order to produce estimates of effects by low-income Constant 0.00967 0.01250 .77 0.4519
status and perform tests of significance. The model is a LowSES2 0.0000087 0.000076 1.15 0.2706
weighted least squares regression where the weights are LowSES -0.0002180 0.000632 -0.35 0.7315
the inverse variances (see the discussion in C1, Grade -0.0009336 0.000670 -1.39 0.1857
Multivariate Results), adjusted for clustering, calculated for R-squared 0.7395 Mean dep. var 0.0137
each effect size as described in Appendix A. The Adjusted R-squared 0.6837 S.D. dependent var 0.0124
dependent variable, ESTOT, is the average effect size, S.E. of regression 0.13985 F-statistic 13.25
also described earlier. There were too few observations to Sum squared resid 0.2738 Prob (F-statistic) 0.0002
model math and reading scores separately.
Joint Significance Tests F-statistic Prob.
The independent variables are percentages of students LowSES and LowSES2 8.31 0.0042
receiving free or reduced lunch (LowSES) and the grade at LowSES and LowSES2 and Grade 13.25 0.0002
which class size was reduced (Grade). The results of a
linear model using LowSES and Grade to predict ESTOT,
summarized in Exhibit C.5, show that LowSES has a
significant positive effect on ESTOT (p=.0015) even after Exhibit C.7
controlling for grade level. Change in Acheivement From Reducing Class
Size: Predicted by Student Low-Income Status
A scatter plot of ESTOT and LowSES (Exhibit 4) indicates
a curvilinear relationship, so a squared LowSES term was .045
added to the model. The results of this model are .040
summarized in Exhibit C.6. Individually, the LowSES and .035
2
LowSES covariates are not statistically significant. Tested .030
Effect Size

jointly, however, they are significant at p=0.0042. Exhibit .025

C.7 illustrates the effect sizes predicted by the curvilinear .020
model (in black) and the linear model (in blue), while .015
holding grade level constant at 3.79 (the mean value for .010
grade level in the model). Actual values are represented in .005
red. While there is no compelling reason to pick one model .000
over the other, they point towards the same general -.005
conclusion, that, all things equal, class size reductions are 5 25 45 65
more effective for classes comprising low-income students
Percent of Low-Income Students
than for other students. Using the curvilinear results, the
effect on achievement of a one-student reduction in class
size for a class with 40 percent low-income students is
more than double that of a class with 20 percent low-
income students (.011 and .005 respectively)

26
 C2. Full-Day vs. Half-Day Kindergarten
This exhibit provides more information on the effect sizes
for full-day kindergarten at various follow-up periods.
Exhibit C.8
Effect Sizes At the End of Kindergarten and
In The Early Years of Education
(Maximum inverse variance weight set at 500)
Follow-up
Grade 2 Grade 4
K Grade 1 or 3 or 5
Effect size 0.1805 0.0108 0.0477 0.0004
Standard Error on Effect
Size 0.02874 0.0375 0.0300 0.0310
Upper limit with 95%
confidence 0.2370 0.0840 0.1060 0.0600
Lower limit with 95%
confidence 0.1242 -0.0628 -0.0111 -0.0612
Number of Effect Sizes 17 6 5
Endnotes
1 For example, in 2000 Washington voters passed Initiative 728 (72
ESSB 6386 §607(15), Chapter 372, Laws of 2006.
2
Washington Learns Steering Committee. (2006). Washington
Learns: World class, learner-focused, seamless education.
Olympia: Washington Learns, pp. 7-8.
3
See: (a) S. Aos, M. Miller, & E. Drake. (2006). Evidence-Based
Public Policy Options to Reduce Future Prison Construction,
Criminal Justice Costs, and Crime Rates. Olympia: Washington
State Institute for Public Policy; (b) S. Aos, M. Miller, & E. Drake.
(2006). Evidence-based adult corrections programs. Olympia:
Washington State Institute for Public Policy; (c) S. Aos, J. Mayfield,
M. Miller, & W. Yen. (2006). Evidence-based treatment of alcohol,
drug, and mental health disorders: Potential benefits, costs, and
fiscal impacts for Washington State. Olympia: Washington State
Institute for Public Policy; (d) S. Aos, R. Lieb, J. Mayfield, M. Miller,
& A. Pennucci. (2004). Benefits and costs of prevention and early
intervention programs for youth. Olympia: Washington State
Institute for Public Policy; and (e) S. Aos, P. Phipps, R. Barnoski, &
R. Lieb. (2001). The comparative costs and benefits of programs to
reduce crime. Olympia: Washington State Institute for Public
Policy.
4
See: F. Cunha, J. Heckman, L. Lochner, & D. Masterov. (2005).
Interpreting the evidence on life cycle skill formation. In Handbook
of the Economics of Education, E. Hanushek & F. Welch (Eds.),
North-Holland. See also: W. Riddell. (2006). The impact of
education on economic and social outcomes: An overview of
recent advances in economics. University of British Columbia:
Department of Economics.
5
As described in the appendix, we calculate mean-difference effect
sizes for each study and then meta-analyze these individual effect
sizes to produce an average effect size for a group of studies on a
particular topic. In general, we follow the procedures in M.W.
Lipsey & D. Wilson. (2001). Practical meta-analysis. Thousand
Oaks: Sage Publications. Many studies of education topics,
however, are based on data that are organized hierarchically:
students are nested in classes, classes are nested in schools, and
schools are nested in districts. To account for this, we adjust effect
sizes and inverse variance weights using methods suggested in
L.V. Hedges. (2006). Effect sizes in cluster-randomized designs.
Institute for Policy Research, Northwestern University, Working
Paper Series manuscript.
6
The disagreements between Hanushek and Krueger over the
effectiveness of policies can be seen in: L. Mishel & R. Rothstein
(Eds.). (2002). The class size debate. Washington DC: Economic
Policy Institute. On the other hand, the agreements between these
two economists on how to calculate benefits of any statistically
significant effect can be seen in: A. Krueger. (2003). Economic
considerations and class size. The Economic Journal, 113: F34-
F64; and E.A. Hanushek. (2004). The economic value of improving
local schools, downloaded from:
<http://edpro.stanford.edu/Hanushek/admin/pages/files/uploads/Ec
onomic%20Value.cleveland%20fed.pdf>.
7
See: Hanushek (2004), citing the work of R.J. Murnane, J.B. Willett,
Y. Duhaldeborde, & J.H. Tyler. (2000). How important are the
cognitive skills of teenagers in predicting subsequent earnings?
Journal of Policy Analysis and Management, 19(4): 547-568. See
also: J. Currie & D. Thomas. (2001). Early test scores, school quality
and SES: Longrun effects on wage and employment outcomes.
Research in Labor Economics, 20: 103-132.
8
For a review of this literature see: W.C. Riddell. (2006). The
4 impact of education on economic and social outcomes: An
overview of recent advances in economics. University of British
Columbia: Department of Economics.
9
For a review of the issues in the debate, see: L. Mishel & R.
Rothstein (Eds.). (2002). The class size debate. Washington DC:
Economic Policy Institute. See also the opposing arguments in: R.
Greenwald, L.V. Hedges, & R.D. Laine. (1996). The effect of school
resources on student achievement. Review of Educational
Research, 66(3): 361-396. E.A. Hanushek. (1996). A more
complete picture of school resource policies. Review of Educational
Research, 66(3): 361-396.
10
percent yes vote) authorizing additional funding for reduced class
size, extended learning programs, educator professional
development, and facility improvements.
11
For a summary, see: R. Ehrenberg, D.J. Brewer, A. Gamoran, &
J.D. Willms. (2001). Class size and student achievement.
Psychological Science, 2(1): 1-30.
12
A.B. Krueger. (1999). Experimental estimates of education
production functions. Quarterly Journal of Economics, 114(2): 497-
532. A.B. Krueger & D.M. Whitmore. (2001). The effect of attending a
small class in the early grades on college-test taking and middle
school test results: Evidence from Project STAR. Economic Journal,
111(468): 1-28.
13
For a study questioning some of the STAR findings, see: E.A.
Hanushek. (1999). Some findings from an independent investigation
of the Tennessee STAR experiment and from other investigations of
class size effects. Educational Evaluation and Policy Analysis, 21(2):
143-168.
14
The two studies that analyzed the same dataset for STAR are:
A. Krueger. (1999). Experimental estimates of education
production functions. Quarterly Journal of Economics, 114(2: 497-
532; and C. Kang. (2005). Effects of small classes on academic
achievement: Evidence from new entrants to Project STAR.
Singapore: National University of Singapore, Department of
Economics.
15
For small changes greater than a one-unit drop in class size, our
effect sizes can be multiplied by the class size change to
approximate the total effect. To date, only a few analysts have
tried to estimate non-linear relationships between class size and
test scores; see Borland et al. (2005).
16
The average return and the range were computed from the
simulation model. The range was set to include 1.5 standard
deviations above and below the mean level of returns in the 5,000
case simulation.
17
The long-term nominal rate of return on the S&P 500 is about
7.4 percent per year. After adjusting for inflation, the real rate is
about 4.4 percent.
18
M. Oelerich. (1984). Should kindergarten children attend school
all day every day? The Journal of the College of Education, Fall:
13-16 (ERIC No. ED254318).
19
Ibid.
27
20
K. Kauerz. (2005). Full day kindergarten: A study of state
policies in the United States. Denver, CO: Education Commission
of the States.
21
Ibid.
22
Preliminary findings from the 2007 Washington State full-day
kindergarten survey. Personal communication with Debra Williams-
Appleton, Office of Superintendent of Public Instruction, March 2,
2007.
23
Complete information on the ECLS-K is available on the website
of National Center for Education Statistics:
<http://nces.ed.gov/ecls/kindergarten.asp>.
24
V.E. Lee, & D.T. Burkam. (2002). Inequality at the starting gate:
Social background differences in achievement as children begin
school. Washington, DC: Economic Policy Institute.
25
Because not all studies provided information on demographics,
we are unable to split the studies into advantaged vs.
disadvantaged children.
26
More detailed information about these effect sizes is provided in
Appendix C2.
27
These groups are derived from several papers. Head Start:
Nielsen & Cooper-Martin (2002) analyzed results separately for
children who attended the local Head Start program the year
before kindergarten. In the Head Start group, 92 percent belonged
to racial or ethnic minorities and 83 percent qualified for free or
reduced lunch. Minority status: DeCicca (2006) analyzed results
through the end of first grade for Black and Hispanic children
separately. Poverty: Cannon, et al. (2006) analyzed results
through third grade for children with family income below the
federal poverty level. Saam & Nowak (2005) analyzed results at
third grade for children eligible for free lunch. DeCicca and
Cannon, et al. used data from the Early Childhood Longitudinal
Study-Kindergarten (ECLS-K).
28
We follow the meta-analytic methods described in M.W. Lipsey, &
D. Wilson. (2001). Practical meta-analysis. Thousand Oaks: Sage
Publications.
29
D. Webbink (2005). Causal effects in education. Journal of
Economic Surveys, 19(4): 535-560.
30
See: Identifying and implementing education practices supported
by rigorous evidence: A user friendly guide (2003, December)
Coalition for Evidence-Based Policy, U.S. Department of
Education, Institute of Education Sciences, National Center for
Education Evaluation and Regional Assistance. Available at:
<http://www.evidencebasedpolicy.org/docs/Identifying_and_Imple
menting_Educational_Practices.pdf>.
31
Lipsey & Wilson (2001), Table B10, equation 22, p. 200.
32
Ibid., Table B10, equation 1, p. 198>.
33
L.V. Hedges. (1981). Distribution theory for Glass’s estimator of
effect size and related estimators. Journal of Educational Statistics, 6:
107-128.
34
Lipsey & Wilson (2001), equation 3.22, p. 49.
35
These formulas are taken from: L.V. Hedges. (2007). Effect
sizes in cluster-randomized designs. Manuscript downloaded from
the author’s website, cited with permission of the author.
36
Lipsey & Wilson (2001), equation 3.23, p. 49.
37
Ibid., equation 3.24, p. 49.
38
Ibid., p. 114.
39
Ibid.
40
Ibid.
41
Ibid., p. 116.
42
Ibid., p. 134.
43
See, for example: A. Krueger. (2003). Economic considerations and
class size. The Economic Journal, 113: F34-F64; and E.A. Hanushek.
(2004). The economic value of improving local schools, downloaded
from <http://edpro.stanford.edu/Hanushek/admin/pages/files/
uploads/Economic%20Value.cleveland%20fed.pdf>.
44
W.C. Riddell. (2006). The impact of education on economic and
social outcomes: An overview of recent advances in economics.
University of British Columbia: Department of Economics.
28
45
Current Population Survey data downloaded from the U.S.
Census Bureau site with the DataFerrett extraction utility:
<http://www.bls.census.gov/cps/cpsmain.htm>.
46
Washington State Economic and Revenue Forecast Council:
<http://www.erfc.wa.gov/pubs/nov06pub.pdf>.
47
United State Bureau of Labor Statistics, Employment Cost Index,
March 14, 2006 release, data for December 2005:
<http://www.bls.gov/news.release/ecec.toc.htm>.
48
See Congressional Budget Office data for the June 2006 report,
Table W-5, at: <http://www.cbo.gov/ftpdocs/72xx/doc7289/06-14-
SupplementalData.xls>.
49
Our estimated return and standard error of the return are calculated
from the findings in: R.J. Murnane, J.B. Willett, Y. Duhaldeborde, &
J.H. Tyler. (2000). How important are the cognitive skills of teenagers
in predicting subsequent earnings? Journal of Policy Analysis and
Management, 19(4): 547-568. This study and others are also
reviewed in: E.A. Hanushek. (2004). The economic value of
improving local schools, pg 6. downloaded from:
<http://edpro.stanford.edu/Hanushek/admin/pages/files/uploads/Econ
omic%20Value.cleveland%20fed.pdf>.
50
Hanushek (2004), p. 7.
51
B. Wolfe & R. Haveman. (2002). Social and nonmarket benefits
from education in an advanced economy. “Proceedings from the
Federal Reserve Bank of Boston's 47th economic conference,”
Education in the 21st Century: Meeting the Challenges of a
Changing World, accessed from: <http://www.bos.frb.org/
economic/conf/conf47/index.htm>. See also a collection of articles
on the topic published in J. Behrman & N. Stacey (Eds.). (1997). The
social benefits of education. Ann Arbor: The University of Michigan
Press. See also: W.C. Riddell. (2006). The impact of education on
economic and social outcomes: An overview of recent advances in
economics. University of British Columbia: Department of
Economics.
52
Office of Management and Budget, Circular A-94 (revised 1992).
53
See Congressional Budget Office report: <http://www.cbo.gov/
ftpdocs/72xx/doc7289/06-14-LongTermProjections.pdf>.
54
For a general discussion of discount rates for applied public
benefit-cost analyses, see: C. Bazelon & K. Smetters. (1999).
Discounting inside the Washington D.C. Beltway. Journal of
Economic Perspectives, 13(4): 213-28. See also: H. Kohyama.
(2006). Selecting discount rates for budgetary purposes, Briefing
Paper No. 29. <http://www.law.harvard.edu/faculty/hjackson/
DiscountRates_29.pdf>.
55
B. Nye, L. Hedges, & S. Konstantopoulos. (1999). The long-term
effects of small classes: A five-year follow-up of the Tennessee
class size experiment. Educational Evaluation and Policy Analysis,
21(2): 127-142.
56
B. Nye, L. Hedges, & S. Konstantopoulos. (2001). The long-term
effects of small classes in early grades: Lasting benefits in
mathematics achievement at grade 9. The Journal of Experimental
Education, 69(3): 245-257.
57
A. Krueger, & D. Whitmore. (2001). The effect of attending a
small class in the early grades on college-test taking and middle
school test results: Evidence from project star. The Economic
Journal, 111(January): 1-28.
58
J. Finn, S. Gerber, & J. Boyd-Zaharias. (2005). Small classes in
the early grades, academic achievement, and graduating from high
school. Journal of Educational Psychology, 97(2): 214-223.
59
Krueger (1999).
Document No. 07-03-2201